graph_trainer: add named EP overlap schedules

[sanketpurandare]

Stack from ghstack (oldest at bottom):

• -> graph_trainer: add named EP overlap schedules #4022
• 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

Add an explicit named-schedule mode alongside the dependency-driven auto
scheduler. Schedules use typed token-exchange, module-FQN, and ready-filler
anchors, and the DeepSeek-v3 policy represents the calibrated forward,
saved-backward, and full-recompute ordering.

Token anchors schedule only the dependency closure needed to launch an
exchange. Every exchange must be covered exactly, and malformed or
dependency-invalid schedules fail with region context.

Why

Named schedules make communication/compute placement reproducible and allow
model-aware alternatives to greedy filling without adding model-specific
branches to the generic scheduler.

Test Plan

• Validate unknown schedule names and DeepSeek-v3 provider applicability.
• Run an exact two-rank DeepSeek-v3 comparison between named and auto graph
  schedules.
• Compare both loss and gradient norm for 20 steps with full recomputation,
  forced-balanced routing, and regional Inductor.

[sanketpurandare]

Closing this experimental named-schedule direction and removing it from the landing stack. The retained stack uses the shared automatic scheduler and bounded MinimalAsyncEP copies.