Clean Inductor complication cache during dynamo dashboard run (#87246)
Implement improvement from https://github.com/pytorch/torchdynamo/issues/1644.
Tested by running `python benchmarks/dynamo/runner.py --print_run_commands --training` and inspecting the generated `run.sh` file for the `--cold_start_latency` flag, e.g.
```
python benchmarks/dynamo/torchbench.py --performance --float32 -dcuda --output=benchmark_logs/inductor_torchbench_float32_training_cuda_performance.csv --training --inductor --no-skip --dashboard -x fambench_xlmr -x detectron2_fasterrcnn_r_50_c4 -x detectron2_fasterrcnn_r_50_dc5 -x detectron2_maskrcnn_r_101_fpn -x detectron2_maskrcnn_r_50_fpn -x detectron2_fasterrcnn_r_50_fpn -x detectron2_maskrcnn -x detectron2_fasterrcnn_r_101_dc5 -x opacus_cifar10 -x detectron2_maskrcnn_r_101_c4 -x pyhpc_turbulent_kinetic_energy -x maml -x detectron2_fasterrcnn_r_101_fpn -x pyhpc_equation_of_state -x detectron2_fasterrcnn_r_101_c4 -x pyhpc_isoneutral_mixing --cold_start_latency
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/87246
Approved by: https://github.com/anijain2305, https://github.com/jansel
