Remove the test retry policy A/B experiment from unstable (#194779)

Removes the retry-policy A/B added in #193469, now that it has run for ten days and answered its question. The leaner policy (no in-process reruns, one process retry) is a small loss on CUDA -- 23 red shards vs 14, 6 experiment-only red commits vs 1 control-only -- and a no-op on CPU, while saving nothing measurable in wall-clock (~15 min across 5,204 CUDA job-hours on main). Full results: pytorch/test-infra#8407.

`PYTORCH_NUM_PROCESS_RETRIES` is kept in `run_test.py` -- it defaults to the pre-experiment value of 2, and lowering it to 1 is a separate free win the experiment surfaced (all 46 control-arm process-retry rescues landed on the first fresh retry).

Authored with assistance from Claude (Claude Code).

Pull Request resolved: https://github.com/pytorch/pytorch/pull/194779
Approved by: https://github.com/pytorchgreenlight