Test what main's coverage misses, and exclude what no input reaches - #656

Stacked on #596, below #651. Tests for lines on main that Codecov showed as never run, and
# pragma: no cover (or no branch) for the ones no input can reach. The library changes
are pragmas and comments only, plus one stale "BRANCH NOT COVERED" comment removed: no
behavior changes here. The bugs these tests turned up are fixed in their own PRs lower in
the stack, or queued (below).

What the new tests cover:

• Monoids and aggregators (test_op.py, test_pickle.py): a bool field's min and max
  identities; a date or text field, which has none, named in the monoid's KeyError, and
  the aggregator not working with the UDT; a user monoid without an identity; a named
  monoid unpickled after its name is gone, registered again with its identity and
  is_idempotent; agg.sum[udt].jit_c_name following its monoid.
• IndexBinaryOp (test_indexbinary.py, test_ssjit.py): no theta binds 0, through the
  typed op or the op given a dtype; C and GraphBLAS scalar thetas; a UDT-typed op given a
  value that is not one; an op with no theta bound refused as an operator by reduce; the
  errors before SuiteSparse:GraphBLAS 9.4; gb.indexbinary.ss absent on the vanilla
  backend.
• The JIT compiler repair (test_ssjit.py, against a stand-in config):
  fix_jit_config(probe=False), which enables the JIT without a test compile; no
  CONDA_PREFIX, so the compiler, flags and libraries come from sysconfig, or None when
  Python records no compiler; a baked-in -isysroot that does not exist, with and without
  xcrun; the import-time repair on a SuiteSparse with no JIT options; an error the JIT
  control does not explain re-raised as it came.
• SuiteSparse extras: dtypes.ss.register_new refusing a bad name or typedef before
  any compile; the element count check of array settings in gb.ss.config (replacing a
  commented-out check on memory_pool, which is no longer an option);
  Matrix.ss.serialize of a UDT whose name is too long for the blob; deserializing a blob
  whose integer sizes SuiteSparse cannot read; gb.ss.diag of a Vector without options;
  the shape of a full array-UDT import without nrows and ncols; test_ss_pack_hyperhash
  with 1025 rows, so SuiteSparse computes the hash on every version checked (8.2.1, 10.5.0
  and 10.5.1), where with 2 rows the pack arm ran on some CI draws only.
• Elsewhere: Matrix Market I/O through a stand-in fast_matrix_market (CI does not
  install it, so that engine never ran); an infix ewise_add of two infix expressions;
  unary.identity and unary.one on a UDT without Numba; the text and HTML reprs of an apply
  naming its UDT scalar operand.

Each pragma carries a reason word, (safety) for unreachable by construction,
(defensive), (unused), (only on failure) or (child process), and a one-line comment
where the reason is not already written beside it. In the library: a UDT's C typedef
asked for without its numpy type, the explicit layout of a dtype that is neither a record
nor an array, get_typed_op's except arm, the typed op finalizer's shutdown guards,
IndexBinaryOp guards that its builder already enforces, _scalar_identity with a name
other than min or max, the int64 config arms (no option has that type since memory_pool
went), a deserialize status check that the read before it already made, and lookups that
every supported build satisfies. In the tests: code that runs only in a spawned child
process (the cross-process pickle tests), on failure, or on a re-run in the same process.
Review tried to reach each excluded line; the four it could reach are tested or counted
again instead.

Bugs found, all on main too:

• (v & v) & w with sizes that differ raised TypeError: _dummy() missing 1 required positional argument instead of DimensionMismatch, also for | and @: fixed in
  Fix Scalar infix expressions with a literal operand #654.
• monoid.min and monoid.max on a record with a timedelta64 field raised ValueError from
  np.iinfo, also from udt in agg.min; a UDT monoid whose op returns another type raised
  a DomainMismatch with no message; Monoid.register_new(lazy=True) dropped
  is_idempotent; and a user op named plus, times, min or max got that built-in monoid's
  generated identity, which gave wrong results silently: all four fixed in Extend the monoid guard to built-ins that carry a UDT fallback func #619.
• A UDF that Numba types for complex values but cannot lower for them (max(x, y), which
  uses >) raised Numba's NotImplementedError at registration, and from op["FC64"] and
  "FC64" in op with is_udt=True: fixed in Learn built-in UDF return types by typing alone, defer the compiles #612.
• An IndexBinaryOp called with no theta bound a BOOL op, so INT64 inputs were cast to bool
  (op() gave [2] where op(0) gave [8]): fixed in Type UDT ops like built-in ops; fix UDT literals, casts and comparisons #591, where op(theta) is typed when
  it runs.
• Numba refused array UDTs of bool, datetime64, timedelta64 or str with a raw
  AttributeError at registration: Type UDT ops like built-in ops; fix UDT literals, casts and comparisons #591 raises a ValueError that names the array and what
  to use instead.

Queued, not fixed in this stack:

• min and max identities for a timedelta64 field (after Extend the monoid guard to built-ins that carry a UDT fallback func #619 it raises the KeyError a
  datetime64 field does).

The same survey found why most of test_ssjit.py's JIT tests skip in CI. The compiler is
fine on Linux and macOS; psg's wheels build GraphBLAS without the JIT, and on Windows the
conda repair did not look where conda installs compilers. #636 now runs test_ssjit.py with
--runslow in the Linux and macOS notebooks jobs (and installs the psg version a wheel or
source draw picks, which had always been the newest), and #652 finds a
conda compiler on Windows.

Coverage, measured locally (suitesparse backend, Numba, --runslow, the full suite) before
review added three more tests: missed lines went from 926 to 836 and partial branches from
546 to 495, 95.90% to 96.32%. Most of what is left is covered only on other CI draws (older
SuiteSparse, vanilla, no Numba) or is #591's own code.

Every commit that adds tests was checked by breaking the code under test: 24 of 27 such
changes made the new test fail, and the other three changed nothing observable; the three
tests added in review each fail with their code broken. The touched
test files pass in random order with fixed seeds in eight configurations (including no
Numba on both backends, CI's C-scalar pass, numpy 2.2, SuiteSparse:GraphBLAS 8.2 and Python
3.15). The full suite passes with --runslow, without Numba on both backends, and in the
C-scalar pass; on Python 3.15, numpy 2.2 and SuiteSparse:GraphBLAS 8.2 the only failure is
test_version, from the local environments.

A CI job that always compiles the JIT (added in review, at Erik's request). The JIT tests
used to run only on random draws that got the suitesparse backend, --runslow, Numba and a
GraphBLAS that compiles, so Codecov's total moved by up to 0.7 points between runs.
Add a CI job that always compiles the JIT, so its coverage stops swinging adds one fixed
entry, (ubuntu-latest, pytest_jit): Python 3.15 and the newest psg from conda-forge, Numba
always, and the full suite with --runslow. Fail the JIT tests instead of skipping them under
GRAPHBLAS_REQUIRE_JIT makes that job turn red when it cannot compile, where it would otherwise
skip the JIT tests and lose their coverage. Every other job skips as before.