[Triton/Gluon] Support caller-defined padding cache slot in fused MLA writer

[tanth47]

Summary

Add an optional pad_slot_id to fused_qk_rope_cat_and_cache_mla so callers can reserve a non-negative cache slot for padding without changing the default AITER behavior.

The argument defaults to -1, so slot 0 remains writable for existing callers. When a caller opts in, both the decode/mixed and extra-prefill KV-write branches skip the selected slot. The behavior is implemented consistently in the Triton and gfx1250 Gluon kernels.

Motivation

Some runtimes reserve a physical KV-cache slot while using padding rows in slot_mapping. The generic SGLang MLA writers already protect this invariant in sgl-project/sglang#36003, but the fused AITER writer bypasses those writers and can still overwrite the reserved slot.

This change keeps AITER generic: no slot is reserved by default, and callers explicitly select one when needed. Downstream context is tracked in sgl-project/sglang#36207.

Tests

Added focused coverage that verifies:

• default behavior still writes slot 0;
• pad_slot_id=0 preserves slot 0 in the decode/mixed branch;
• pad_slot_id=0 preserves slot 0 in the extra-prefill branch;
• ordinary positive slots are still written.

The new value is a compile-time kernel argument, so the default predicate can be constant-folded to the existing behavior.

Local checks: Python compilation, Black, Ruff, and git diff --check. GPU execution is left to AITER CI.

[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 5010 --add-label <label>

PR title tags:
Component tags ([Triton/Gluon], [HIP], [CK], [ASM], ...) are added to the PR title automatically from the changed files and re-synced on every push — change-type tags like [fix]/[Perf] and op tags like [MLA] are left untouched. Add the no-auto-title label to opt this PR out of title tagging.

[tanth47]

@k50112113 Could you take a look and approve the fork workflows? This adds an opt-in padding-slot guard to the fused MLA writer while preserving the existing default. It also updates the gfx1250 Gluon mirror; please add ci:gfx1250-ffm-triton if that suite is appropriate (contributors cannot apply the label).