perf(flydsl): Kimi-K3 B1 route-sort + prepared-MoE handoff

[JohnQinAMD]

Purpose

Re-cut of the Kimi-K3 B1 route/prepared-MoE work onto current main, DCO-signed.

This is the AITER half of the campaign's largest single accepted step (+20.3%
end-to-end, 50.635 → 60.909 tok/s/GPU at TP8/B1/8K-in/1K-out; route boundary 1.62x).
The original PR (#7) carried a nine-commit stack that mixed this
work with the MLA split override, the KDA decode/f_b fusion, the latent MoE tail and
the B1 gate projection. Those four have since been re-cut and published separately
(ROCm#4405, ROCm#4495, ROCm#4496, and #6), and #7 was closed CONFLICTING —
leaving this part, the actual route/handoff change, with no live PR anywhere.

This branch is only the four commits that were never re-cut:

commit	change
[Kimi-K3] Fuse gfx950 B1 routing and MoE sort metadata	fused B1 route-sort kernel + kimi_k3_moe_route entry point
perf(moe): add Kimi-K3 B1 stage1 direct tile	direct-tile stage1 selection + tuned config row
perf(moe): use atomic Kimi-K3 B1 stage2 reduction	atomic stage2 reduction + tuned config row
perf(flydsl): consume Kimi-K3 prepared MoE routing	kimi_k3_moe_handoff — accept prepared routing from the caller

+1212/-2 across 10 files, mostly new files under aiter/ops/flydsl/.

Test Plan

• four op_tests/flydsl_tests/ suites ship with the change — route-sort, B1 stage1,
  B1 stage2, and the MoE handoff
• the campaign gate this was accepted under: endpoint-measured TP8 / batch 1 /
  8192-in / 1024-out, temperature 0, --ignore-eos, with a paired GSM8K-100 accuracy
  check against the parent build

Test Result

Measured at the time of acceptance, on 8x MI355X (gfx950), against its exact parent:

	parent	with this change
decode	50.635 tok/s/GPU	60.909 tok/s/GPU (+20.3%)
TPOT	19.718 ms	16.418 ms
route boundary	—	1.62x

GSM8K 100q 5-shot temp 0: no accuracy regression against the parent.

These are the original campaign numbers, not a re-run of this branch. The change is
byte-identical to what was measured — the four commits are cherry-picked unmodified,
and the conflict-free rebase onto current main touched none of them — but the endpoint
result itself has not been reproduced on today's main.

[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 29 --add-label <label>

[github-actions]

⚠️ [ruff] <B008> _{reported by reviewdog 🐶}
Do not perform function call fx.Stream in argument defaults; instead, perform the call within the function, or read the default from a module-level singleton variable

[github-actions]

⚠️ [ruff] <F821> _{reported by reviewdog 🐶}
Undefined name Dict

[github-actions]

⚠️ [ruff] <F821> _{reported by reviewdog 🐶}
Undefined name Dict

[github-actions]

⚠️ [ruff] <I001> _{reported by reviewdog 🐶}
Import block is un-sorted or un-formatted

Suggested change

	import importlib
	import math
	from dataclasses import replace
	import aiter
	import pytest
	import torch
	from aiter import dtypes
	from aiter.fused_moe import fused_moe_2stages, moe_sorting
	from aiter.jit.utils.chip_info import get_gfx_runtime
	from aiter.ops.flydsl.kimi_k3_moe_handoff import (
	KimiK3MXFP4ExpertRequest,
	KimiK3MXFP4W13Layout,
	kimi_k3_mxfp4_expert_handoff,
	kimi_k3_mxfp4_expert_mode,
	supports_kimi_k3_mxfp4_expert_handoff,
	)
	from aiter.ops.flydsl.moe_common import GateMode
	pytestmark = pytest.mark.skipif(
	import importlib
	import math
	from dataclasses import replace
	import pytest
	import torch
	import aiter
	from aiter import dtypes
	from aiter.fused_moe import fused_moe_2stages, moe_sorting
	from aiter.jit.utils.chip_info import get_gfx_runtime
	from aiter.ops.flydsl.kimi_k3_moe_handoff import (
	KimiK3MXFP4ExpertRequest,
	KimiK3MXFP4W13Layout,
	kimi_k3_mxfp4_expert_handoff,
	kimi_k3_mxfp4_expert_mode,
	supports_kimi_k3_mxfp4_expert_handoff,
	)
	from aiter.ops.flydsl.moe_common import GateMode
	pytestmark = pytest.mark.skipif(