(WIP) perf(kimi-k3): tune small-M MoE decode

[panditsa]

Summary

• extend the joint latent/shared decode pipeline from M=1 through M=4
• rebalance routing and W2 work for M=2/4
• tune TP-local A16W4 tiles and split top-k route work across CTAs

Performance

4K/1K TP8/EP8 CUDA graphs, median of three:

M	Prior equivalent stack tok/s	Head tok/s	Change	Prior TPOT	Head TPOT
2	84.77	93.39	+10.2%	22.864 ms	20.663 ms
4	152.23	156.34	+2.70%	24.954 ms	24.255 ms

The comparison uses the last equivalent EP8 boundary: intervening #1137/#1138 only add unselected A8W4 foundations and #1139 registers the exact TP8/EP1 shape. The measured EP8 execution path is unchanged until this PR.

Tests

• BF16 SiTU M=1/2/3/4 coverage: 4 passed
• router dispatch/per-token FP32 route coverage: 2 passed
• fused RMSNorm + projection M=1/2/3/4 graph coverage: 4 passed
• all 18 measured 4K/1K requests completed with valid lengths

Stack

• depends on (WIP) feat(moe): select gfx950 TP A8W4 SiTU #1139

[jhinpan]

Exact #1132 screen of the only M=8-applicable change here (the sigmoid-router warp-count line), applied independently to current main (9196fa27) with forced acceptance=1 at 131072/1024 TP8/EP8:

• baseline: 36.80 ms TPOT, 19.7647 tok/s
• router candidate: 36.77 ms TPOT, 19.7753 tok/s

That is only ~+0.05% throughput / 0.03 ms and is not an issue-level material win. The larger W13/W2/pipeline changes remain gated to M<=4 or the TP-local EP1 shape, so I will not extract this line into a separate #1132 PR.