Support EP sentinels

[IlyasMoutawwakil]

This PR adds support for EP through sentinel experts (experts_ids >= E) #50

[IlyasMoutawwakil]

@GarlGuo we have some pretty good numbers with this PR

compared to grouped_mm, we noticed that sonicmoe scales pretty well with naive EP (all tokens are passed but only a subset is routed on each rank, the non-routed ones get a sentinel value and are skipped by the grouped gemm computation).
this naive EP approach allows us to shard experts and reduce compute but not token hidden state memory (all-to-all would reduce both), the nice thing about it is that it doesn't add any cpu-gpu syncs (because the shapes are not routing dependent) so torch.compile/fullgraph/cuda-graph compatible

[tridao]

woah this scaling is pretty good!