feat: Allow loading and serializing with tensorizer

[sangstar]

This draft PR aims to unintrusively integrate tensorizer in to exllamav2's model loading machinery, along with adding tests for correctness and avoiding regressions.

Currently in a draft PR stage. The tests have been most recently updated, but the core logic needs to be made less intrusive and less smelly.

Still to add:

• Add comments to test file for better clarity
• Make way tensorizer is exposed in config machinery less intrusive
• Decide whether to use io_handler for all Tensorizer I/O stuff or retire it altogether
• Consider rethinking the way packaging tensorizer configurable stuff is done, whether it needs a dedicated arguments class or if just packing them in their config class is appropriate
• Decide if .state_dict should be a public attribute
• Allow passing TensorDeserializer args to calls to TensorDeserializer with some wrapper