[Compression] Update sparsity calculation lifecycle when fetching the compressor

[dsikka]

Summary

• Updating the lifecycle based on feedback from research and meta
• Very slow saving of models due to irrelevant stats computation #1293 is addressed @eldarkurtic

1. No longer always infer sparsity

• If there is existing sparsity in the model that should be accounted for while compressing the model, the user is required to indicate this by providing a sparsity config. This is not a common case and primarily something we've seen from research. As a result, it does not need to be the default pathway
• If no sparsity config is provided, only check for sparsity if a recipe with sparsity was applied OR the user explicitly would like to apply a sparsity compressor by setting skip_compression_stats parameter
• Based on these changes, we no longer use the global_sparsity to dictate whether or not a sparse compressor is applied

2. Change the name of skip_compression_stats to skip_sparsity_compression_stats and set to True by default

3. Rename sparsity_config to sparsity_metadata_config for clarity

4. Add additional log outputs to make it a lot more clear as to what steps are running

5. Fix log outputs to prevent cutting off

Follow-ups:

• Need to update passing saving arguments properly - they are currently not being passed apart from save_compressed and there is no clear API on what the arguments are/how they should be passed, both to save_pretrained and when providing an output dir to oneshot

Testing

• Tested a sparse-only model
• Tested a model with both sparsity and quantization
• Tested a model with existing sparsity
• Tested a model with only quantization

For all cases, the model compresses to the expected format.

[github-actions]

👋 Hi! Thank you for contributing to llm-compressor. Please add the ready label when the PR is ready for review.

Note: This is required to complete the testing suite, please only add the label once the PR is code complete and local testing has been performed.

[rahul-tuli]

Nice, the functionality seems to be all there! LGTM pending a small nit and some clean up (docstrings and such)