Add a WASM SIMD128 8x4 SGEMM microkernel

[jjerphan]

Summary

Follow-up to #5983 / #5986 / #5680 / #4023: replace the 4×4 SGEMM microkernel for WASM128_GENERIC with an 8×4 tile so adjacent f32x4 A loads match the V8 WasmRevecReducer shape. Double stays on the existing 4×4 path.

• SGEMM_DEFAULT_UNROLL_M/N become 8/4. SGEMMKERNEL uses kernel/wasm/gemmkernel_8x4_wasm128.c with gemm_{n,t}copy_8 for A and gemm_{n,t}copy_4 for B.
• STRMMKERNEL switches to generic trmmkernel_8x4.c so TRMM unroll matches SGEMM. DTRMMKERNEL stays trmmkernel_4x4_wasm128.c; DGEMM packing and unroll stay 4×4.
• Relaxed madd is used in the SGEMM inner k-loop when compiled with -mrelaxed-simd / WASM_RELAXED_SIMD=1 (same as GEMM in Add a 4x4 WASM SIMD128 GEMM microkernel #5983). Remainders keep 4/2/1 scalar (and a 4-wide SIMD panel for bm & 4).

Benchmarks

Node / Emscripten, TARGET=WASM128_GENERIC, USE_THREAD=0, COMMON_OPT=-O2, WASM_RELAXED_SIMD=1. Same machine, pinned to one P-core (taskset -c 0), median of timed samples. Speedup > 1 means this branch is faster than develop.

Geomean vs develop (n = 64…4096): SGEMM 1.52x, SSYRK 1.47x, STRMM 1.30x, STRSM 1.22x. DGEMM/DSYRK/DTRMM/DTRSM unchanged within noise (~1.00x).

op	n	develop MFLOPS	8x4 MFLOPS	speedup
sgemm	64	26947.2	37264.9	1.383
sgemm	128	27032.8	42076.1	1.556
sgemm	256	28086.9	43561.5	1.551
sgemm	512	28842.5	41872.3	1.452
sgemm	1024	28300.8	45321.0	1.601
sgemm	2048	29265.3	45075.0	1.540
sgemm	4096	30399.6	47028.1	1.547
ssyrk	64	25290.1	31863.6	1.260
ssyrk	128	27646.8	39643.7	1.434
ssyrk	256	30235.2	44223.4	1.463
ssyrk	512	30996.6	46699.1	1.507
ssyrk	1024	30669.0	47689.7	1.555
ssyrk	2048	31164.8	48452.6	1.555
ssyrk	4096	31017.6	48258.7	1.556
strmm	64	25397.0	28038.7	1.104
strmm	128	30082.0	33779.7	1.123
strmm	256	31072.7	37505.8	1.207
strmm	512	31762.2	42779.3	1.347
strmm	1024	31704.7	43052.4	1.358
strmm	2048	31370.6	47616.9	1.518
strmm	4096	31151.7	47383.5	1.521
strsm	64	16118.3	13295.8	0.825
strsm	128	21350.9	21217.3	0.994
strsm	256	25570.2	29843.4	1.167
strsm	512	28219.7	37165.9	1.317
strsm	1024	30017.9	42520.5	1.417
strsm	2048	30836.5	46110.6	1.495
strsm	4096	30776.0	46643.2	1.516

Test plan

• Node CBLAS odd-size GEMM/TRSM/TRMM check
• utest under node (106/106)
• CBLAS ctest L1/L2/L3 (x{s,d,c,z}cblat{1,2,3}) under node, including SGEMM/DGEMM, STRMM/DTRMM, STRSM/DTRSM, SSYRK/DSYRK

[jjerphan]

Local Node / Emscripten correctness gate on this branch (TARGET=WASM128_GENERIC, USE_THREAD=0, WASM_RELAXED_SIMD=1):

• CBLAS odd-size GEMM/TRSM/TRMM check: OK
• utest: 106/106 ok
• CBLAS ctest L1/L2/L3 (x{s,d,c,z}cblat{1,2,3}): no computational failures

[jjerphan]

I wonder whether this rather be changed in another PR.

[martin-frbg]

GEMM is usually linked to TRMM (and SYMM) in OpenBLAS, so I'm fine with having it here

[jjerphan]

I have opened #6024 which should be valuable for properly testing all those new kernels for WASM.