Add PEFT benchmarking script in thunder/benchmarks

[riccardofelluga]

What does this PR do?

This PR picks up and replaces #1978 with some tweaks. It introduces PEFT benchmarks for single gpu and multi-gpu with fsdp2 as deistributed strategy.

PR review

Can be tested by running:

python thunder/benchmarks/benchmark_peft.py --model deepseek-ai/DeepSeek-R1-Distill-Qwen-7B --trust-remote-code --attn-implementation sdpa --max-steps 10 --mbs 1 --seq-length 4096 --compile thunder --fixed-num-hidden-layers 2

Multi-GPU:

torchrun --nproc_per_node=2 thunder/benchmarks/benchmark_peft.py --model deepseek-ai/DeepSeek-R1-Distill-Qwen-7B --trust-remote-code --attn-implementation sdpa --max-steps 10 --mbs 1 --seq-length 4096 --compile thunder --fixed-num-hidden-layers 2

Side notes

I've extended the logging from the original PR and removed the prints. Now this uses the log levels to print more or less information depending on --verbose flag.

@crcrpar in the fsdp2 part I carried it over from the benchmark_litgpt.py file, but I've commented out the litgpt specific layers. I've left it there for now but I am planning to remove them after the review, let me know if you think I should find the equivalent layers in transformers and adapt that code.

cc. @wprazuch

[KaelanDt]

Great add! Left some small comments

[t-vi]

Thank you @riccardofelluga @KaelanDt

At some point we might look at splitting out the various requirements here, but not now.