Add MoEDispatcher for expert-parallel token dispatch/combine

[mawad-amd]

Summary

• Promote MoE dispatch/combine from examples/31_expert_sharded_moe/ to production iris.ccl module
• Add MoEDispatcher class with pre-allocated buffers and handle-based API (dispatch() → combine())
• Add moe_utils.py with routing helpers (ExptAssignment, RaggedTensorMetadata, topk, reduce)
• Add ctx.ccl.moe_dispatcher() factory method
• Add benchmark comparing iris vs naive all_to_all dispatch

Key Design

• Direct iris.store() scatter — no AllToAll/AllToAllv; sparse per-token routing via symmetric heap
• Pre-allocated buffers in __init__ — dispatch, combine, all-gather routing buffers allocated once
• Handle-based API — dispatch() returns opaque DispatchHandle consumed by combine()
• Reuses proven kernels — _convert_dp_to_ep and _convert_ep_to_dp are identical to the example

Performance (iris vs naive all_to_all)

T_local	iris (us)	naive (us)	Speedup
32	28,462	19,368	0.68x
64	24,177	34,645	1.43x
128	24,757	66,385	2.68x
256	24,679	125,387	5.08x
512	25,987	227,935	8.77x

iris shows near-constant latency, winning at T_local >= 48 tokens/rank with up to 8.8x speedup.

Full performance study: https://gist.github.com/mawad-amd/c43c02e3662a180a88e26faa31c62f52

Test plan

• MoE example tests pass (2/2)
• New MoE dispatch tests pass (21/21) — e2e, dispatch-only, combine-only, buffer-reuse, topk=1, handle-frozen
• CCL regression tests pass (no new failures)
• Parametrized: T_local=[32,128], H=[64,256], k=[1,2], dtype=[bf16,fp32]

🤖 Generated with Claude Code