(WIP) perf(kimi3): join TP MoE reductions

[panditsa]

Summary

• enable joint routed/shared reduction for native TP as well as EP layouts
• acquire producer-direct outputs from the TP-EP reduction group
• pass activated shared inputs through the tuned M<=4 A8W4 expert kernel

Performance

Exact parent a1a0f8e7 → head 30cac366, 4K/1K TP8/EP1 CUDA graphs, median of three:

M	Parent tok/s	Head tok/s	Change	Parent TPOT	Head TPOT
2	97.17	103.64	+6.66%	19.939 ms	18.621 ms
4	175.34	185.64	+5.88%	21.693 ms	20.380 ms

Each run had one first-repetition TTFT disturbance; the two steady repetitions agreed closely and the reported median rejects that startup effect.

Tests

• TP8 execution policy and reduction-group coverage pass
• joint projection/output reuse is covered for M=1/2/3/4/8
• targeted suite: 6 passed, 28 deselected
• all 18 measured requests per boundary completed with valid 4096-token prompts and 1024-token outputs

Stack

• depends on (WIP) perf(comm): extend fused AttnRes through M16 #1142