qsmla: adapt Shared-B TransB usage to pto-spec #257 (compile fix for TileOP-API >= 748ac69)

[CYR-Firework]

Summary

TileOP-API #104 (01d62c4) redefined the cooperative Shared-B TransB control as a physical-storage-shape declaration per pto-spec #257 (ASL BundleMatrixSharedBPrimarySchemaLegal):

TransB	physical Shared source
0	[N, K] (K contiguous)
1	[K, N] (N contiguous)

The QSMLA kernel still used the legacy math-transpose-flag reading, so since TileOP-API 748ac69 it fails the compile-time effective-K / cooperative-D shape static_asserts (template_asm.hpp:4219/4232) — QSMLA is outside the microbenchmark guard coverage, so this was not caught by the daily reports.

This PR adapts the 5 call sites in quant_sparse_flash_mla.hpp. Both K and V are staged as SharedMatrixRight<kvdtype, kTk, kTd> (one KV token per row, D contiguous), so:

• QK (Q @ K^T): the physical [kTk, kTd] layout IS the [N, K] form for the K^T operand → drop transpose_b() (TransB=0) — 4 sites
• PV (P @ V): the same physical layout is the [K, N] form for the V operand (N contiguous) → add transpose_b() (TransB=1) — 1 site

Note the transposition itself does not disappear — under the new contract it is carried implicitly by the physical-layout declaration (TransB=0 ⇒ the engine transposes on read), which is why the numeric results below are bitwise identical.

Verification

Environment: TileOP-API 64ef554 + SuperScalarModel fix/gfrun-shared-tmatmul-layout-257 @ 6a60fb5 (the model-side TransB adaptation de3ea07f) + LLVM 10abd6e+, baseline tag ops-20260914 + this patch.

gfrun 4-PE (-s softcore.multiThreadNum=4, HIF8 in / BF16 out, atol=rtol=2e-2 vs CPU golden):

Case	Result	max_abs
swa_small	32768/32768 (100%)	0.004920
csa_small	32768/32768 (100%)	0.004677
ori_sparse_small	32768/32768 (100%)	0.009071

max_abs is bitwise identical across every baseline before and after the #257 contract change (ops-20260908 → ops-20260914), confirming the adaptation direction.

Compile: 0 errors / 0 warnings for all three IMPLs (swa_tadd_4pe / csa_tadd_4pe / ori_sparse_tadd_4pe, QSMLA_DTYPE=HIF8, embedded inputs).

References

• Contract source: Clarify TMATMUL Shared-tile matrix layouts and transpose semantics pto-spec#255, Specify Shared TMATMUL physical layout semantics pto-spec#257
• TileOP-API compile-side adaptation: Sync TMATMUL Shared-B effective K/N shape checks with PTO-ISA #255 LinxISA/Linx-TileOP-API#103 → [FA] FA算子性能优化：提升CUBE利用率并减少冗余向量操作 #104, [gfrun][NA] Vector Numeric Validation Failures #105
• Ecosystem breakage report (QSMLA double-blocked): QSMLA blocked by the #104/#257 TransB contract change: kernels fail to compile with legacy usage, and #257-adapted usage is rejected by SuperScalarModel gfrun (opposite TransB semantics) LinxISA/Linx-TileOP-API#122
• Model-side fix (companion of this PR): SuperScalarModel fix/gfrun-shared-tmatmul-layout-257 @ 6a60fb5, commit de3ea07f ("fix(gfrun): adapt Shared TMATMUL layout")

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.

[CYR-Firework]

Superseded by #151, which consolidates this work (commit-for-commit, rebased on current main) together with the single-PE CSA kernel. Closing in favor of #151.