[Quantization Format] Add functionality to infer format

[dsikka]

Summary

• Move format infer functionality from llmcompressor to compressed-tensors - [QuantizationFormat] Remove code inferring format vllm-project/llm-compressor#1786
• Simplifies functionality to set the format based on the quant_scheme and sparsity scheme, if a format is not already set by the user
• Use the per module format to determine the global format for the model
• Adds dosctrings
• Add loguru
• Return dense if no format inferred, not None

Next Steps

• We can set the per module format when the quantization scheme is initialized with some work on how we're handling the sparsity config / structure. Will work on simplifying this next to allow this. Otherwise, we currently do it during compression time at the end

[brian-dellabetta]

LGTM. Few code-related things, nothing pops out as wrong in implementation but deferring to others on that 😄

[shanjiaz]

Looks good to me!

[kylesayrs]

Can you add a comment on QuantizationScheme which specifies that None means that the value will be inferred by infer_and_set_per_module_quantization_format before compression?

Otherwise looks great, thanks for doing this

[kylesayrs]

#449