[torch-bisect] fix repro cmdline bug and enhance with more cases - #172

Last good commit: 9d49044
First bad commit: 34cdf49
Metric: geomean_speedup
Threshold: 0.95
Rel_tol: 0.001
Dtype: amp
Arch: b200
Device: cuda
Suite: timm_models
Compiler: cudagraphs
Mode: training
Branch: main
Repo: pytorch/pytorch
Model: inception_v3
Regression: 3.59x -> 3.33x (7.5%)
REPRO_CMD: python benchmarks/dynamo/timm_models.py --performance --amp --training --only inception_v3 --inductor

No regression detected: current value 3.881, 3.888 / 3.881 - 1 == 0.1803658850811677%, threshold 4.0%)

Between commits 7ba94e7f (Apr 12) and 437c81a2 (Apr 13), 20+ torchbench models dropped 13-22% speedup on H100 bfloat16 inference:
nanogpt: 8.67x to 6.72x (-22%)
phlippe_resnet: 5.63x to 4.51x (-20%)
squeezenet1_1: 4.20x to 3.39x (-19%)
mobilenet_v3_large: 5.80x to 4.78x (-18%)

python benchmarks/dynamo/torchbench.py --performance --bfloat16 --inference --only mobilenet_v3_large --inductor

python benchmarks/dynamo/torchbench.py --performance --bfloat16 --inference --only nanogpt --inductor