graph_trainer: unify chunked grad collective normalization

[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

Normalize duplicate chunk-local parameter-gradient collective chains after
either eager or graph chunking. Compatible chunk gradients are accumulated
before a single all-reduce or reduce-scatter, and any wait, view, or dtype-cast
suffix is replayed once on the normalized result.

Why

Dense FSDP and EP schedulers should see the same gradient buckets regardless
of how the two chunks were produced. A shared post-chunking pass keeps that
contract in one place and avoids producer-specific scheduling behavior.

Test Plan

• Trace real SimpleFSDP eager and graph chunk paths in replicate and
  fully-sharded modes.
• Verify eager duplicate collectives normalize and graph collectives remain
  canonical.
• Exercise suffix replay through a naturally produced mixed-precision cast.