Adds a matrix_sum aggregate that collects per-row matrix handles and
performs a single LAGraph_Matrix_Sum (extract-tuples + one GrB_Matrix_build)
at finalize, as an alternative to matrix_agg's serial pairwise eWiseAdd.

Gated behind WITH_MATRIX_SUM=1 / -DONESPARSE_HAVE_LAGRAPH_MATRIX_SUM since it
requires the LAGraph matrix-sum branch (GraphBLAS/LAGraph#407); default
builds against stock LAGraph are unaffected. Dockerfile-debug/test.sh take
LAGRAPH_REPO/LAGRAPH_REF build args to build against the PR branch.

Includes plus_binaryop() helper and benchmark scripts under bench/.

On a zero-overlap .grb corpus, matrix_sum scales O(N) vs matrix_agg's O(N^2)
(6.8x at 64 matrices, 39x at 512).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Add a third matrix-summing aggregate alongside matrix_agg (serial pairwise
eWiseAdd) and matrix_sum (single concat + GrB_Matrix_build). matrix_binary_sum
has the same interface and produces identical results to matrix_sum, but uses
LAGraph_Matrix_Binary_Sum (a binary reduction tree of eWiseAdds) for a
different parallel/time/space tradeoff.

- src/matrix/matrix_binary_sum.c: gated behind ONESPARSE_HAVE_LAGRAPH_MATRIX_SUM
  (same LAGraph PR #407 branch as matrix_sum). internal-state sfunc collects
  dup'd handles with a reset-callback for cleanup; finalfunc calls
  LAGraph_Matrix_Binary_Sum. Default dup PLUS, optional binaryop arg.
- bench/matrix_sum.sql: register 1-arg and 2-arg matrix_binary_sum aggregates.
- bench/matrix_sum_bench.sql, bench/matrix_sum_grb_bench.sql: assert all three
  engines agree and compare their timings.
- Dockerfile-debug, test.sh: add LAGRAPH_CACHE_BUST build arg to force a fresh
  LAGraph clone when a branch gains new commits under an unchanged ref (the
  cached clone otherwise lacked the new symbol).

Verified: builds with no warnings, all 33 regression tests pass, and
matrix_agg = matrix_sum = matrix_binary_sum exactly (default PLUS and MAX dup).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The matrix-combining matrix_agg(a matrix) folded each row into a running
accumulator with a single serial GrB_eWiseAdd, which is ~O(N^2) on
low-overlap data (it re-touches the whole growing accumulator on every
row). On the graphchallenge corpus it took 477s at K=512 vs 12s for
matrix_sum, so it occupied no useful corner of the time/space curve.

Reimplement it as a streaming binary-counter merge: a carry array levels[r]
holds the sum of exactly 2^r already-seen matrices; each input enters at
rank 0 and is carried up, summing-and-freeing occupied levels, until it
lands in an empty slot. The occupied levels are folded at finalize. This is
O(N log N) and frees each input the moment it is absorbed, so it never holds
all N inputs at once.

The aggregate moves from STYPE=matrix to STYPE=internal with a dedicated
finalfunc (matrix_agg_combine_final). The per-type element builder
matrix_agg(i, j, v) and its matrix_agg_final are unchanged. The default dup
operator changes from TIMES to PLUS, matching matrix_sum and
matrix_binary_sum, so matrix_agg(a) == matrix_sum(a) == matrix_binary_sum(a).
It is pure GrB_eWiseAdd, so it ships ungated in the stock extension.

Benchmark (graphchallenge low-overlap corpus): matrix_agg is now ~15x faster
at K=512 and is the only one of the three engines to reach K=4096 (it is the
most memory-frugal); matrix_sum OOMs at 4096 and matrix_binary_sum at 1024.
Results are bit-identical where the engines overlap.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Adds onesparse.nthreads (GxB_GLOBAL_NTHREADS), onesparse.chunk
(GxB_GLOBAL_CHUNK), onesparse.print_1based (GxB_PRINT_1BASED), and
onesparse.hyper_switch (GxB_HYPER_SWITCH) as PGC_USERSET GUCs so users
can tune SuiteSparse parallelism and format thresholds per-session.

Also adds PGC_POSTMASTER GUCs for all writable JIT fields:
onesparse.jit_cache_path, onesparse.jit_c_compiler_name,
onesparse.jit_c_compiler_flags, onesparse.jit_c_linker_flags,
onesparse.jit_c_libraries, onesparse.jit_c_preface,
onesparse.jit_error_log, onesparse.jit_use_cmake.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>