[inductor] Parallelize Max Autotune step 2: Use multiple GPUs (#109127)
Test Plan:
`python test/inductor/test_max_autotune.py`
`TORCHINDUCTOR_AUTOTUNE_IN_SUBPROC=1 TORCHINDUCTOR_MAX_AUTOTUNE=1 python benchmarks/dynamo/torchbench.py --device cuda --performance --backend inductor --inference --only hf_Bart`
`TORCHINDUCTOR_AUTOTUNE_MULTI_DEVICE=1 TORCHINDUCTOR_AUTOTUNE_IN_SUBPROC=1 TORCHINDUCTOR_MAX_AUTOTUNE=1 python benchmarks/dynamo/torchbench.py --device cuda --performance --backend inductor --inference --only hf_Bart`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/109127
Approved by: https://github.com/shunting314, https://github.com/eellison
ghstack dependencies: #109126
