Drop unify's ignored scalar flags inside the library, and share the UDT leaf walkers - #665

Stacked on #663. Follow-ups to #591 from review. This is a refactor with no change in behavior. Every existing
test passes unchanged, and one test is added.

unify's scalar flags. dtypes.unify has ignored is_left_scalar and is_right_scalar
since #591, which types a literal before it reaches unify. The library no longer passes
them:

• the fill-value and ewise_union calls in Vector, Matrix and Scalar;
• get_typed_op (in graphblas.core), which took them only to pass them on and no longer
  takes them.

unify keeps both keywords, and its docstring now says it accepts and ignores them.

Walkers.

• _has_half_leaf, _has_narrow_float_leaf, _has_duration_leaf,
  _numba_untyped_array_element and _raw_bytes_element now test the leaves of one walker,
  dtypes._iter_leaves. It walks records, nested records, arrays of records and arrays of
  arrays, as each of them did.
• _check_leaves_compare uses _leaf_dtypes, which its inner function repeated.
• _int_beyond_leaf is gone. _beyond_as_infinity read a literal by the same rules. It now
  returns None when no int is beyond its leaf's range, so a comparison walks the literal
  once.
• _check_udt_pair and _corresponding_leaves ask _record_leaf_mismatch whether two records
  pair: the same nesting, the same names in the same order, and the same shape for each field.
  Each words its reason as before. Array UDTs keep their own rules, which differ on purpose:
  operands broadcast, but a stored element does not.

Some walkers stay separate. Each reads a literal by its own rules, and merging them would
change a result:

• _refuse_misfit_fields: a non-sequence has no parts.
• _weak_literal_udt's record reader: only a number fills a record, and a wrong length
  refuses the whole literal.
• _nats_in_leaf_units: a list is a sequence of elements.
• _as_record_parts: it turns a dict into a tuple, and keeps a scalar as it is.
• _leaf_dtypes: it treats an array of records as one leaf.

Tests. A new row in test_jit_udt_np_type_numba_cannot_type checks a field that is an
array of arrays of bool, which numpy keeps unflattened. It still gets the message for an array
Numba cannot type.

A differential check compared the old and new walkers on 362 dtypes:

• every scalar kind, including text, bytes, raw void, dates, durations, float16 and complex;
• arrays of them, and arrays of arrays;
• random records with array fields;
• nested records and arrays of records;
• titled fields and out-of-order offsets.

It also covered 8,014 literal values and 131,044 dtype pairs. All 665,406 results matched on
numpy 1.24.4, 2.1.3, 2.4.6 and 2.5.4. The full suite passes with --runslow on both backends,
without Numba, and with the CI's C-scalar substitution.

Questions for Erik

🤖 Generated with Claude Code