mla: add fp8 qh32 seqlen=1 persistent kernel support on gfx950

[alexioslyrakis-amd]

Summary

Adds the mla_a8w8_qh32_qseqlen1_gqaratio32_ps kernel for gfx950 (MI350X), covering the decode case with gqa_ratio=32, fp8 Q/KV, and seqlen_q=1.

• asm_mla.cu: add seqlen=1 dispatch branch for gqa_ratio=32 fp8/fp8 (sub_Q=32); update error message to list all supported seqlens (1/2/4)
• v1_2_device.cuh: add seqlen=1 to natively_supported conditions for gfx950 nhead=32 fp8
• mla.py: add gfx950/nhead=32/fp8/seqlen=1 to the native-path selector in mla_decode_fwd
• mla_asm.csv: register new .co entry for the qh32 seqlen=1 persistent kernel
• mla_a8w8_qh32_qseqlen1_gqaratio32_ps.co: compiled kernel binary
• test_mla.py, test_mla_persistent.py: enable nhead=32 fp8 seqlen=1 test paths

Test plan

• python op_tests/test_mla_persistent.py --nhead 32,1 --dtype fp8 --kv_dtype fp8 --batchSize 512 --ctxLen 4096 on MI350X

[github-actions]

🏷️ CI Guide

Runs automatically on every PR:

• ✅ Pre-checks (submodule verification, code formatting)
• ✅ Aiter op tests (gfx942 + gfx950)
• ✅ Triton tests on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label	Tests
ci:triton-300x	Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
ci:sglang	SGLang integration tests: DeepSeek-R1-MXFP4 accuracy, Qwen 3.5 accuracy
ci:atom	ATOM benchmark: DeepSeek-R1-0528, GPT-OSS-120B
ci:atom_full	ATOM accuracy suite for PR and main models from ATOM models_accuracy.json
ci:vllm	vLLM benchmark: GPT-OSS-120B, DeepSeek-R1-0528, Kimi-K2.5
ci:all	All standard extended tests (excludes ci:atom_full)

Only add ci:atom_full for FlyDSL or Triton upgrades.
Add labels via the sidebar or gh pr edit 3304 --add-label <label>