This PR re-implements the existing verifier, making the following improvements:

• Enabling the verification of more models on inference
  • Previously, it expected the models to return a single output tensor
• Enabling the verification of training
  • There was a plumbing bug in TorchBenchModel.train
  • Running training for a few iterations
• Model cleanup
  • Delete each used model so that we don't run out of memory on large models
• Use PyTorch functions for checking whether the accuracy is acceptable
  • Instead of using MRE, we use the same thing PyTorch does

In order to do so, here's a summary of the changes:

• Introduce BenchmarkModel methods: tolerance(), use_cosine_similarity, and skip_verifier()
  • The logic is taken from PyTorch, which also uses torchbench.yaml
• Introduce force_dtype parameter when loading a model
  • So that we can run models on eager fp64
• More meaningful verification codes
• Move reset_rng_state and cleanup to util.py
• Change how we access the YAML configuration file, replacing the raw strings

cc @miladm @zpcore