minimal_async_ep: enable eager EP overlap

[sanketpurandare]

Stack from ghstack (oldest at bottom):

• graph_trainer: support MinimalAsyncEP graph chunking #3956
• -> minimal_async_ep: enable eager EP overlap #3955
• minimal_async_ep: pipeline bounded high-priority copies #4021
• minimal_async_ep: bound receive buffer capacity #4020
• minimal_async_ep: split token exchange launches and waits #3954
• graph_trainer: unify chunked grad collective normalization #3953

Summary

Enable eager two-chunk EP overlap for MinimalAsyncEP. Allocate independent symmetric-memory buffer sets, give dispatch and combine distinct pools sized to their own receive capacities, select stable buffer ownership while tracing each chunk, and teach the shared EP scheduler to recognize MinimalAsyncEP launch and wait markers.

Force rematerialization inside MinimalAsyncEP MoE regions so receive-buffer views are never saved across backward. The existing one-set path remains unchanged when overlap is disabled.

Why

Launch/wait separation is not sufficient when adjacent chunks can reuse static receive storage. Buffer ownership, separate exchange pools, and rematerialization are part of the correctness contract when communication moves across expert compute. Exact pool sizing also avoids allocating combine storage at the larger dispatch bound.

Test Plan

• Run real two-rank dispatch, expert computation, combine, reduction, and backward through two independent buffer sets.
• Compare outputs and gradients with a local reference.
• Verify dispatch/combine pool capacities and storage separation.
• Verify bounded CTA configuration and prep/copy/wait stream placement.
• Run the two-rank dispatcher, full-Inductor waits, and CUDA graph wrapping in H100 CI.