Support MLA decode with nhead < 16 by transparent pad-to-16

[ChuanLi1101]

Summary

• For MLA models with small query head counts (e.g., Kimi-Linear-48B-A3B with TP=8 giving nhead=4), AITER's ASM kernel has no pre-compiled support for gqa_ratio < 16, causing decode failures.
• This PR adds transparent head padding within AITER: when nhead < 16 and divides 16 evenly, Q is padded to 16 heads via repeat_interleave, the nhead=16 ASM kernel runs, then the output is un-padded.
• Adjusts C++ persistent metadata generation and Python metadata sizing to accept nhead < 16.
• Adds nhead=4 test configurations to test_mla.py and test_mla_persistent.py.

Changes

• aiter/mla.py: Pad Q heads to 16 when nhead < 16 (both non-persistent and persistent paths), un-pad output before return. Add safe entries in get_block_n_fp8.
• csrc/kernels/mla/metadata/v1_2_device.cuh: Add pad_to_qh16 logic for nhead < 16 in persistent metadata generation.
• aiter/ops/attention.py: Relax num_head_qo % 16 assertion in get_mla_metadata_info_v1 to allow nhead < 16 with effective_num_head = 16.
• op_tests/test_mla.py: Add nhead=(4,1) to default test configurations.
• op_tests/test_mla_persistent.py: Add nhead=(4,1) to default test configurations.

Test plan

• nhead=4 BF16 decode test on MI355X (gfx950): all checkAllclose passed
• nhead=16 BF16 regression test: all passed, no regressions
• nhead=4 FP8 decode test (future work - needs ASM kernel support)
• CI pipeline validation

[github-actions]

🏷️ CI Guide

Runs automatically on every PR:

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

Extended tests (opt-in via labels):

Label	Tests
ci:triton-355	Run Triton tests on MI355 in addition to MI325
ci:sglang	SGLang integration tests
ci:atom	ATOM benchmark (DeepSeek-R1 + GPT-OSS)
ci:vllm	vLLM benchmark
ci:all	All of the above

Add labels via the sidebar or gh pr edit 2577 --add-label <label>

[ChuanLi1101]

cc @valarLip @carlushuang — Requesting expedited review. This enables MLA decode with nhead < 16 (required for GLM-5 TP=8 on MI355X). This is blocking vLLM-side PRs (vllm-project/vllm#36855, vllm-project/vllm#38665) for customer-facing GLM-5 inference. See also #2563 for the upstream issue.