[benchmark] add option to enable CompiledAutograd

[crcrpar]

What does this PR do?

CompiledAutograd seems to speed up FSDP2 and I checked it with torchtitan.
I however somehow do not find it beneficial for litgpt models.

setting: pjnl-20241209, 8H100

torchtitan Llama-3-8B

This uses activation checkpointing, since the provided config by default uses it -- https://github.com/pytorch/torchtitan/blob/05a8b5e4c1de979c4b49ff36e6b09d6055db29b1/train_configs/llama3_8b.toml#L53-L55

CompiledAutograd	Performance (tps)	Memory (GB)
N	6244	51.2
Y	7200	43.0

litgpt llama-2-7b-hf

CompiledAutograd	Performance (tokens/s/GPU)	Memory (GB)
N	11722.76	39.13
Y	10702.33	52.61