perf(flydsl): fuse Kimi-K3 KDA decode + f_b projection

[JohnQinAMD]

Summary

Two commits: (1) fuse the Kimi-K3 KDA decode path (causal conv + recurrent
update + gated RMSNorm) into one gfx950 kernel; (2) fold the 128x128 f_b_proj
into the KDA decode state core (head-local, in-workgroup).

Effect

• Removes one low-occupancy GEMV + the g1 global materialization + a dependency
  bubble per KDA layer.
• Complete-chain 1.36x (9.76 -> 7.15 us B1); banked endpoint 50.063 tok/s.

Provenance

Banked change #3 (see perf_changelog.md). Independent PR off main.

AI assistance (OpenAI Codex) was used; a human submitter must review every line
before merge.

Made with Cursor

[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 5 --add-label <label>

[JohnQinAMD]

Superseded by #16, rebuilt as the same two reviewable decode-core and incremental f_b commits on current main. #16 resolves only the FlyDSL export drift, passes Ruff/diff checks, restores DCO/disclosure on both commits, and remains draft for fork Black/GPU CI.