Fix Dynamo Benchmarks: Stabilization iters counted in latency (#196060)
Summary:
## Summary
Fix HF LLM compilation-latency accounting by measuring matched eager and compiled generations, while keeping additional compiled calls as untimed cache stabilization. Peak-memory and Dynamo-stat collection remain unchanged.
There was a bug in the Dynamo benchmarks that included stabilization iterations in the compilation latency calculation. `run_performance_test()` sets `niters=1` for HuggingFace LLMs. Eager uses `niters` for warmup explicitly. However, compiled warmup skips `niters` and uses `warmups()`'s default of 5.
`compilation_time = dynamo_latency - eager_latency`, so this compares five compiled geneartions againts one eager generation.
## Test Plan
Adds a focused model-free regression test, and I manually ran a Qwen3-0.6B check with 1000 input and 2000 output tokens
| Accounting | Compilation latency |
|---|---:|
| Before: 5 compiled calls - 1 eager call | 44.565 s |
| After: 1 compiled call - 1 eager call | 44.448 s |
| Overcount removed | 0.117 s (0.26%) |
X-link: https://github.com/pytorch/pytorch/pull/196060
Approved by: https://github.com/anijain2305
Reviewed By: huydhn
Differential Revision: D118929058
fbshipit-source-id: c0295d223500ae1b757a7ddda06f8ea01990cdaf