(WIP) perf(kimi3): enable batched AttnRes runtime fusion

[panditsa]

Summary

• allow the backend-selected fused AttnRes combine for supported multi-token decode batches
• hoist next-layer MLP/attention partials whenever the projection kernel reports support
• join the producer stream before gfx950 M<16 reductions that consume AttnRes scratch

Validation result

This WIP is not ready to land: activating the batched fusion regresses end-to-end decode even though the isolated kernel in #1142 is faster.

4K/1K TP8/EP1 CUDA graphs, median of three:

M	Equivalent parent tok/s	Head tok/s	Change	Parent TPOT	Head TPOT
2	103.64	93.99	-9.3%	18.621 ms	20.308 ms
4	185.64	172.06	-7.3%	20.380 ms	21.459 ms

The parent figures are the exact #1144 head. #1145 only selects its sharded tail at M>=8, so M=2/4 are unchanged at the #1146 parent boundary.

The likely issue is orchestration rather than fused-kernel latency: joining the producer stream and consuming AttnRes scratch removes overlap that the prior split path retained. Runtime activation should remain gated until the dependency/stream schedule is redesigned and reprofiled.

Tests

• batched runtime-dispatch test passes and confirms the fused Iris combine is selected when supported
• (WIP) perf(comm): extend fused AttnRes through M16 #1142 covers world-size-8 M=1/2/4/8/16 numerical correctness and CUDA graphs
• all 18 measured head requests completed with valid lengths

Stack

• depends on (WIP) perf(kimi3): shard the TP MoE final projection #1145

WIP blocker

• restore the runtime gate or recover the lost producer/consumer overlap before landing

[jhinpan]

The exact #1132 M=8/EP8 point differs from the published M=2/4 EP1 blocker. On current main (9196fa27), #1141 + #1142 + this runtime activation gives:

• forced 131072/1024 TP8/EP8: 36.25 ms TPOT and 19.9703 tok/s (3 runs), versus 36.80 ms / 19.7647 baseline: -1.49% TPOT, +1.04% throughput
• normal 4K/1K DSpark: 155.078 tok/s versus 152.949 baseline (+1.39%)
• exact token identity versus ordinary greedy passes at 4K/1K and 131K/128 under forced rejection

So the activation is beneficial at M=8/EP8, but still regresses the already-reported M=2/4 EP1 points and only moves the #1132 round ratio from 2.157x to ~2.125x. This supports redesigning/gating the orchestration by shape/topology rather than landing the current activation globally.