Observer Restructure: Remove Observers, calibration, and applying frozen steps from lifecycle

[dsikka]

Summary

• Remove Observers from compressed tensors as well as the init, calibration and frozen steps responsible for initializing, applying, and removing the observers
• Updates the forward pass to no longer explicitly apply calibration; expects hooks to calibrate inputs and outputs as part of the forward pass
• apply_quantization_status will only be able to apply initialize_module_for_quantization and compress_quantized_weights. The FROZEN and CALIBRATION statuses are still valid QuantizationStatus values but compressed tensors will no longer be able run calibration or "apply" frozen statuses (i.e. remove observers)
• Tests have been updated such that they can mock calibration for per token, per tensor, group and channel quantization
• No longer wrap the forward pass for attention layers as this was done purely for calibration; the kv cache object has been moved to llm-compressor

Once landed, we'll need to land llm-compressor updates to add in Observers right after. Both PRs should be reviewed together: https://github.com/vllm-project/llm-compressor/pull/837/files#diff-b9472770e3291ffd5f7e0adbc074a5b2f83ca00e8dc937e230dcc365f4e2f954

Testing

• Tested w4a16, quantized kv cache, and int8 w8a8 with the llm-compressor changes

[kylesayrs]

Overall looks like a clean cut, I'd like to do a little more testing before I approve

[rahul-tuli]

This PR makes me happy!