Skip rebuilding FX forward after post-lowering and before adjacency split.

[micwill755]

Summary

• Skip gm.recompile() in the post-lowering flush (default recompile=False) and in fast_partition before split(). Conversion (TRTInterpreter) and the adjacency splitter walk graph.nodes and do not call gm.forward().
• Keep DCE + lint. Callers that execute the module as Python pass post_lowering(..., recompile=True) (converter harness, LLM static-cache). split() still rebuilds forward() when it constructs _run_on_acc_* / _run_on_gpu_* GraphModules.
• Klein (3364-node lowered graph): prelude reprint was 36 ms; after skip, fast_partition wall is 118 ms (was ~155–171 ms). split() still reprints the full ACC child twice (~37 ms each) plus an expensive 9-node wrapper reprint (~36 ms).

PR 8 of the compile-startup stack; based on #4598.

Test plan

• tests/py/dynamo/lowering/test_skip_conversion_recompile.py (post_lowering does not recompile; DCE still runs; Interpreter still correct; recompile=True rebuilds forward; fast_partition does not recompile the input module)
• tests/py/dynamo/models/test_lowering_cache.py (roundtrip still executes after explicit recompile of the in-memory module)
• tests/py/dynamo/partitioning/test_000_fast_partitioning.py
• Klein lowered graph: prelude recompile 36 ms; fast_partition wall 118 ms after skip

---

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