Transforms benchmarks experiment

[brian-dellabetta]

This PR leverages the pre-built tasks LLMCompresorTask and LMEvalTask to create an example script to run several llm-compressor recipes at a time, in a single pipeline, in parallel. Each model is then run through an evaluation using lm_eval. Results available at https://spaces.redhat.com/spaces/vLLM/pages/714211727/Transforms+Benchmarks+v1

• I had to make a few updates, pinning dependency versions so that they would install correctly in clearml, based on feedback from @Chibukach
• I added some logic to make sure the LMEvalTask installs vllm with VLLM_USE_PRECOMIPLED=1. This was needed to run an experiment with a transforms feature on main that hadn't been released yet. I think we want to always do this, but I can move it to an input to the LMEvalTask constructor if that is preferred.