[Attention] Expose paged MQA SplitKV override

[AMD-yanfeiwang]

Summary

• expose an optional SplitKV override on deepgemm_fp8_paged_mqa_logits
• preserve the existing launch heuristic exactly when the override is omitted
• validate explicit values and reject ambiguous use with VarCtxSchedule
• add focused regression coverage for validation, legacy resolution, and explicit pass-through

Motivation

The paged MQA kernel already accepts a runtime split count, but the public wrapper always derives it from TotalCuCount, TileQCount, and WavePerEU. This prevents callers from tuning the split independently for production shapes.

On MI355X (gfx950), a diagnostic three-process sweep for the DSV4 shape rows=6144, width=33280, ChunkK=256 found SplitKV=2 consistently faster than the current automatic value of 10:

Run	Auto 10	Explicit 2	Speedup
1	2582.667 us	2546.246 us	1.01430x
2	2590.807 us	2558.686 us	1.01255x
3	2582.267 us	2547.466 us	1.01366x

Aggregate producer gain was 1.382%; all tested split values passed bitwise correctness probes. This PR exposes the mechanism only. It intentionally does not hard-code a shape policy from a single diagnostic shape.

Compatibility

SplitKV=None executes the exact previous formula, so existing callers and launch behavior are unchanged. The underlying Triton/Gluon kernel already accepts a dynamic split count; no kernel ABI or cache-key change is needed.

Validation

• 13 passed in test_pa_mqa_logits_split_kv.py
• public-wrapper MI355X check: zero bitwise mismatches for explicit SplitKV=2
• Ruff, py_compile, and git diff --check pass

[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:gfx1250-ffm-triton	Run the five-shard gfx1250 FFM Triton test suite
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 4775 --add-label <label>