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

[JohnQinAMD]

Summary

Add a gfx950 Kimi-K3 KDA decode kernel, then incrementally fuse the head-local
128x128 f_b projection into the same workgroup.

The first commit fuses causal convolution, recurrent state update, and gated
RMSNorm. The second removes the low-occupancy f_b GEMV, its global g1
materialization, and the dependency bubble before the state core.

Dispatch and fallback

The public wrapper validates device, shape, dtype, stride, cache layout, batch,
and gate-bound contracts. Unsupported requests are rejected before launch and
remain owned by the vLLM fallback. The kernel is exposed only through the
FlyDSL availability boundary.

Performance

Kimi-K3 TP8 batch-one decode on MI355X:

Chain	Time/layer	Relative
Unfused control	9.76 us	1.00x
Decode + f_b	7.15 us	1.36x

The accepted endpoint moved from 47.644 to 50.063 tok/s (+5.1%) and reached
19.953 ms TPOT.

Correctness and validation

• Focused tests cover contract rejection, cache/state mutation, graph replay,
  and fused-vs-unfused output/state comparison.
• The exact clean AITER source plus the final vLLM perf(flydsl): fuse Kimi-K3 mixed pre-route projections on gfx950 #12-perf(flydsl): select validated Kimi-K3 KDA schedule #14 adapter stack passed
  all 9 focused cases in the immutable MI355X validation image, including both
  canonical convolution-cache layouts.
• The accepted endpoint stack containing this exact kernel passed frozen
  GSM8K first 100 at 100/100 with zero invalid responses.
• Ruff check and Ruff format pass all five changed files.
• git diff --check passes.
• Both commits carry DCO and AI-assistance trailers.

Fork Black and Ruff CI pass. The PR is ready for review; hardware workflows
were triggered by the ready-for-review transition.

Disclosure

Developed with assistance from OpenAI Codex. The author reviewed the kernel
contracts, numerical path, tests, and benchmark evidence.

[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 16 --add-label <label>