[graph_trainer] Add EP overlap eager chunking scaffolding

[sanketpurandare]

Stacked PRs:

• [MoE] Use CPU split-size sum for EP permute output size #3627
• [graph_trainer] Add EP overlap scheduling pass #3328
• [graph_trainer] Add graph EP chunking pass #3325
• ->[graph_trainer] Add EP overlap eager chunking scaffolding #3363
• [graph_trainer] Support hinted symbolic input dims in tracing #3362

---

[graph_trainer] Add EP overlap eager chunking scaffolding

Introduce the public graph_trainer EP-overlap configuration surface and the eager chunking producer for its chunk metadata contract. The new compile options select the logical chunk dimension, the chunking strategy, and one supported module-root pattern: all transformer blocks or all MoE blocks. Sequence chunking is limited to MoE block roots because attention needs full K/V context.

Keep EP overlap validation in configs.py next to the other graph_trainer compile config validation. The eager producer wraps selected module forwards during model parallelization, splits tensor inputs into two chunks, calls the original forward once per chunk, and materializes tensor outputs with cat. It emits the same chunk metadata that graph chunking will emit, so the later scheduling pass can consume either producer through one contract.

Add shared MoE EP region annotations for dispatcher dispatch/combine bodies, config fingerprinting for the new options, and a generic trace-input-preparation hook in GraphTrainer. This commit does not add graph chunking or communication-overlap scheduling.

This pass stack relies on pending PyTorch support for hinted unbacked symbolic dimensions in the tracing and distributed compiler paths:

• FakeTensor folded matmul: [ATen][FakeTensor] Handle unbacked dims in folded matmul pytorch#183397
• ProxyTensor SDPA tracing: [ATen][ProxyTensor] Preserve unbacked batch dims in SDPA tracing pytorch#183398
• Inductor bucketing trace isolation from ambient unbacked symbols: [Inductor][Bucketing] Isolate bucketing traces from ambient unbacked symbols pytorch#183495
• Inductor collective bucketing with hinted unbacked SymInts: [Inductor][Bucketing] Make collective bucketing tolerate hinted unbacked SymInts pytorch#183544
• DTensor sharding padding for hinted even unbacked shards: [DTensor] Use explicit hints for unbacked sharding pytorch#183545
• HOP fake traces with discarded unbacked symbols: [HOP][Dynamic Shapes] Ignore discarded unbacked symbols in fake traces pytorch#183837
• FlexAttention chunked unbacked input extents: [Inductor][HOP] Handle chunked unbacked FlexAttention shapes pytorch#183838
• FakeTensor trace metadata for hinted symbolic storage: [FakeTensor] Add hinted symbolic storage size metadata pytorch#183839
• Inductor symbolic stride ordering with unbacked hints: [Inductor] Handle hinted and fallback unbacked symbols pytorch#183840

Test Plan:

• Covered by the full graph_trainer pass, numerics, and H100 integration test runs after the stacked graph chunking and scheduling commits.