[benchmark] add option to use torchao's float8 of dynamic scaling with fsdp2

[crcrpar]

This adds an option to use float8 of torchao.
An example command to use float8 with FSDP2:

torchrun --nproc-per-node 8 --local-ranks-filter 0 --role rank --tee 3 thunder/benchmarks/benchmark_litgpt.py --model_name Llama-2-7b-hf --compile inductor --distributed_mode fsdp2 --shard_mode zero2 --use_torchao_fp8_linear true --use_torchao_fp8_allgather true --use_torchao_fp8_precompute_scale_for_fsdp true

note that --use_torchao_fp8_precompute_scale_for_fsdp true keeps models in fp32.

Llama-2-7b-hf on 8 H100, using pjnl-20240821

as of ff66203, torchao of pytorch/ao@a0376bf

if compiler is torch, fsdp2 is used.

compiler	executors	bs	Tokens/s/GPU	Memory Used
torch	fp8: linear, all-gather, & precompute	1	14640.35	34.26
torch	fp8: linear, all-gather, & precompute	2	17993.14	49.69
torch	fp8: linear & all-gather	1	14587.93	34.26
torch	fp8: linear & all-gather	2	17912.73	49.70
torch	fp8: linear	1	13897.13	40.64
torch	fp8: linear	2	17219.84	56.12
torch	thunder_inductor_cat_cudnn_dynamo	1	12506.03	40.21
torch	thunder_inductor_cat_cudnn_dynamo	2	13890.66	61.82
thunder	inductor_cat_cudnn_transformerengine	1	14274.20	52.88
thunder	inductor_cat_cudnn_transformerengine	2	16329.36	74.01

[lantiga]

Stamped!