fix: remove force nt path in tensorwise grouped gemm

[RuibinCheung]

Description

The non-NT layout grouped GEMM backends (triton and flydsl) are now performant enough on MI350 , so this workaround is no longer needed. This PR removes it and lets TENSORWISE grouped GEMM take the original code path on all architectures: the backward grad_a is computed directly from the untransposed weight with trans_b=not ctx.trans_b, matching what gfx942 has been doing all along.

This is a pure code-path simplification for the TENSORWISE granularity. The ROWWISE, BLOCKWISE and MX_BLOCKWISE paths are untouched, and the identically named helper in gemm_fp8.py (non-grouped GEMM) is intentionally left in place. Keeping non-NT layout in gemm because we still use hipblaslt backend in gemm. And the non-NT layout performance is not good enough in hipbalslt backend.

Fixes # (issue)

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

Please list the changes introduced in this PR:

• Removed the _deter_use_nt_layout_gemm_in_bwd helper from primus_turbo/pytorch/ops/grouped_gemm_fp8.py, together with the now-unused is_gfx942 import.
• FP8GroupedGemmTensorFunc.forward: dropped the transposed-weight cache branch (both the b_t reuse and the transpose(-1, -2).contiguous() fallback), and now always saves the untransposed quantized_b for backward. The b_t argument is unused and annotated as such, consistent with a_t.
• FP8GroupedGemmTensorFunc.forward: no longer stores use_nt_layout_gemm_in_bwd on ctx.
• FP8GroupedGemmTensorFunc.backward: collapsed the two grad_a branches into the single original call with trans_b=not ctx.trans_b. The grad_b (variable-K wgrad) call is unchanged.

Checklist:

• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[Copilot]

Pull request overview

This PR simplifies the FP8 grouped GEMM (TENSORWISE) autograd path by removing the MI350/MI450-era backward workaround that forced an NT-layout code path. It standardizes backward grad_a computation to always use the original call form (trans_b=not ctx.trans_b) and always saves the untransposed quantized weight for backward.

Changes:

• Removed the architecture-gated _deter_use_nt_layout_gemm_in_bwd helper and the related gfx942 check/import.
• Simplified FP8GroupedGemmTensorFunc.forward to always save quantized_b (dropping the transposed-weight cache branch).
• Simplified FP8GroupedGemmTensorFunc.backward to a single grad_a call path using trans_b=not ctx.trans_b.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.