[CK_TILE] FMHA BWD Optimizations for D48 for GFX950

[DDEle]

Motivation

To optimize HDim=48 cases.

Technical Details

Update to ROCm/composable_kernel@95bdc74
Update to ROCm/composable_kernel@2d1c9e2

Test Plan

MAX_JOBS=$(nproc) pytest op_tests/test_mha.py -v

Test Result

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[Copilot]

Pull Request Overview

This PR updates the composable_kernel submodule to implement FMHA (Fused Multi-Head Attention) backward pass optimizations specifically for D48 configurations on GFX950 hardware.

• Updates composable_kernel submodule commit to include FMHA BWD optimizations
• Targets D48 dimension size optimizations for GFX950 GPU architecture
• Focuses on backward pass performance improvements for attention mechanisms

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[DDEle]

It seems that test_gemm_a8w8_blockscale_mi350 fails (coredump) with a high probability with this CK update, while there is a small probability of coredump in current aiter main branch (with linked CK version).

Another pattern of this failure is that the problem only appears in the first run. The test_gemm_a8w8_blockscale_mi350 runs smoothly in following runs (where jit cache exists).

[valarLip]

Test failed: op_tests/test_mha.py ?