minimal_async_ep: pipeline bounded high-priority copies

[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

Run dispatch preparation on a normal stream, then hand row copies to a high-priority stream. Bound overlapping copies with a persistent CTA grid so expert compute retains execution resources; keep the no-overlap path unbounded because it lies on the critical path. The unbounded kernel preserves its two-dimensional row-tile grid.

Expose the grid as --compile.ep_overlap.minimal_async_ep_num_copy_ctas. The default is the measured 50-CTA GB300 setting, while None selects an unbounded overlap control for hardware-specific tuning.

Why

Moving copies to another stream does not reserve SM capacity for expert work. A bounded persistent launch provides an explicit resource limit while stream priority minimizes gaps between dispatch preparation and row copies. Preserving the tiled unbounded launch avoids serializing its critical-path work.

Test Plan

• Verify CTA-grid propagation into all MinimalAsyncEP dispatchers.
• Reject invalid configured grid sizes.
• Exercise bounded and unbounded launch grids.
• Validate copy correctness and prep/copy stream placement in the two-rank dispatcher test.