Apply Torch-TRT decomps during dynamo.trace so compile skips a second AOT retrace.

[micwill755]

Summary

• Apply Torch-TRT's decomp table during dynamo.trace() (the path torch_tensorrt.compile(nn.Module) already uses) so AOT Autograd runs once, matching Inductor / torch.compile(backend="torch_tensorrt").
• dynamo.compile() skips ExportedProgram.run_decompositions when the exported program stamp matches the current settings. Stock torch.export EPs, or a settings mismatch, still decompose.
• No new user-facing API (export_for_tensorrt, export_and_compile, skip_decompositions).

PR 5 of the compile-startup stack; based on #4540.

Test plan

• tests/py/dynamo/models/test_one_shot_decomp.py (stamp match, skip when stamped, stock export still decomps, settings mismatch re-decomps, inplace add_ decomposed during export, trace() stamps)
• Flux.1-dev transformer export+decomp only: second AOT 14.4s → 0s
• Flux2 Klein 4B NVFP4 cold compile (Victor harness): run_decompositions 2.59s → skip; compile phase 33.7s → 28.8s; steady-state unchanged

---

_{Stack created with GitHub Stacks CLI • Give Feedback 💬}