(WIP) perf(comm): extend fused AttnRes through M16

[panditsa]

Summary

• extend the fused Iris attention all-reduce plus AttnRes epilogue from M=1 to M=1..16
• use per-token double-buffered staging with separate ready and consumed epochs
• keep the public communication wrapper shape-driven

Performance

Isolated fused world-size-8 CUDA-graph latency at head a1a0f8e7:

M	Fused all-reduce + AttnRes
1	13.69 µs
2	14.52 µs
4	16.23 µs
8	17.38 µs
16	21.79 µs

Exact parent → head 4K/1K TP8/EP8 M=2, median of three, is flat: 93.98→93.95 tok/s (-0.04%) and 20.534→20.565 ms TPOT (+0.15%). This is expected at this stack boundary because the Kimi runtime still gates batched AttnRes fusion to M=1. #1146 removes that runtime gate; this PR provides and validates the batched kernel capability.

Tests

• world-size-8 numerical comparison against the split all-reduce + AttnRes reference passes for M=1/2/4/8/16
• each M is exercised repeatedly in eager mode and through CUDA-graph capture/replay
• all 12 measured 4K/1K A/B requests completed with valid lengths

Stack

• depends on (WIP) perf(comm): add two-stage producer-direct reduction #1141
• runtime activation for batched decode is in (WIP) perf(kimi3): enable batched AttnRes runtime fusion #1146

[jhinpan]

Exact #1132 validation of #1141 + this PR + #1146 on current main (9196fa27):

• forced 131072/1024 TP8/EP8 baseline: 36.80 ms TPOT, 19.7647 tok/s
• stack, 3 measured runs: 36.25/36.25/36.25 ms; 19.9702/19.9672/19.9734 tok/s
• mean change: -1.49% TPOT, +1.04% throughput
• normal 4K/1K DSpark: 155.078 tok/s vs current-main 152.949 (+1.39%), so the 95% protection gate passes
• forced outputs remain token-identical to ordinary greedy for 4K/1K repeated text and exact 131K/128 natural text

This is a real but small E2E win: the forced round ratio moves only from 2.157x to about 2.125x, not near the 1.5x target. The full orchestration is required; #1141 standalone could not register with SMG.