perf(flydsl): add Kimi-K3 FP8 latent MoE tail

[JohnQinAMD]

Summary

Add an incremental rowwise OCP-E4M3 prepack and gfx950 batch-one persistent
kernel for the Kimi-K3 latent-MoE tail. The kernel fuses RMSNorm, FP8-weight
GEMV, BF16 materialization, and the shared-expert add.

This PR is stacked on #17. It keeps FP32 RMS/dot-product
accumulation, FP32 per-row scales, and BF16 activation/output, and fails closed
outside the exact contiguous (1, 3584) -> (1, 7168) gfx950 contract.

Correctness

The real-weight rotating-bank campaign checked 24 cases:

• worst RRMSE versus BF16: 0.020959;
• minimum cosine: 0.999780;
• minimum SNR: 33.573 dB;
• worst RRMSE versus dequantized FP8: 1.487e-5.

Tests cover support/quantizer/schedule rejection, both numerical oracles, graph
replay, output reuse, and changed input.

Performance

MI355X/gfx950, 92 rotating real weights, 11 trials, 30 graph replays/trial:

Path	ms/token
BF16 parent	0.920660
FP8 candidate	0.668757
Saving	0.251903

Speedup is 1.376673x. The component is below the 0.300 ms/token standalone
endpoint floor, so it receives no isolated endpoint credit. In the Wave40
bundle, TPOT improved 12.194182 -> 11.886058 ms and throughput 82.006322 ->
84.132180 tok/s/GPU; GSM8K first 100 was 100/100 with zero invalid and zero
paired changes.

Validation

• Ruff check/format pass all three changed files.
• Exact clean current-main GPU suite: 7 passed in the immutable MI355X
  validation image.
• git diff --check passes.
• The commit carries DCO and AI-assistance trailers.

Disclosure

Developed with assistance from OpenAI Codex. The author reviewed the precision
contract, tests, schedule, and performance evidence.