[Model][Perf] Fuse Beam's TP all-reduce with its sandwich norms and MoE finalize
MNNVL one-shot/two-shot all-reduce built on FlashInfer's Lamport
protocol, with Beam's FP32 norm-residual-norm epilogue and an optional
top-k MoE finalize before the reduction. FusedMoEConfig gains
defer_moe_finalize(allow_ep=True) so the finalize can also be deferred
under replicated-token EP (TP=1, EP>1).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Woosuk Kwon <woosuk@inferact.ai>