[benchmarks] Default to functionalization disabled.

[ysiraichi]

This PR defaults benchmarks execution to PyTorch/XLA without the functionalization layer. In summary:

• Add --enable-functionalization command-line argument
• By default, set XLA_DISABLE_FUNCTIONALIZATION=1

Reasoning: after running experiments, it became clear that disabling functionalization reduced overhead, improving performance of, mainly, non-dynamo configurations. The speedup we get from this is:

	Inference	Training
Dynamo	1.11x	1.14x
Non-Dynamo	1.3x	1.44x

---

Note: since functionalization is required for correctly supporting aliasing and in-place mutation on views. Turning this option off means that there will be more model failures. Specifically:

Inference + NonDynamo:

• hf_Longformer: execution failure
• llama: execution failure
• pytorch_stargan: execution failure

Inference + Dynamo:

• hf_Longformer: execution failure
• llama: execution failure
• pyhpc_turbulent_kinetic_energy: verifier failure
• pytorch_stargan: execution failure

Training + NonDynamo:

• hf_Longformer: execution failure
• pytorch_stargan: execution failure

Training + Dynamo:

• hf_Longformer: execution failure
• pytorch_stargan: execution failure

cc @miladm @JackCaoG @zpcore

[zpcore]

Thanks, LGTM, those hard coded tests are really annoying when it comes to adding new option to the benchmarking script.

[zpcore]

Can you fix the benchmark test? Thanks