Add skip_missing_weights_context

[kylesayrs]

Purpose

• Add skip_missing_weights_context which is used when loading a quantized model whose state dict does not align model definition weights, ie quantized models (run_compressed or otherwise)

Changes

• Remove init skipping from save_pretrained_wrapper. Init skipping is only useful when loading models, not saving
• Remove transformer logging skipping from tests into skip_missing_weights_context, which is more reusable

Testing

• TODO show example script
• TODO investigate using from custom_offload_map

[github-actions]

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