[Observer Restructure]: Remove MemoryLess Observer; use helper function for dynamic quantization

[dsikka]

Summary

• As part of the observer restructure, this is the first of the PRs to move observers out of compressed-tensors
• Removes the MemoryLessObserver and replaces it with a simple helper function which can determine the zp/scales dynamically during the forward pass
• Will now return None for observers when running dynamic quantization and update the config such that null shows up in the config

Testing

• Tested llm-compressor examples for int8 and fp8, w8a8 for both; will no longer attach observers and the config will look like the example below:

"quantization_config": {
    "config_groups": {
      "group_0": {
        "input_activations": {
          "actorder": null,
          "block_structure": null,
          "dynamic": true,
          "group_size": null,
          "num_bits": 8,
          "observer": null,
          "observer_kwargs": {},
          "strategy": "token",
          "symmetric": true,
          "type": "float"
        },
        "output_activations": null,
        "targets": [
          "Linear"
        ],
        "weights": {
          "actorder": null,
          "block_structure": null,
          "dynamic": false,
          "group_size": null,
          "num_bits": 8,
          "observer": "minmax",
          "observer_kwargs": {},
          "strategy": "channel",
          "symmetric": true,
          "type": "float"
        }
      }
    },

• Tested compatibility with existing models (can still be loaded and run as expected)

[mgoin]

Need to update observer: str = Field( to observer: Optional[str] = Field( in src/compressed_tensors/quantization/quant_args.py

[kylesayrs]

lgtm, would like to see more validation/error catching happen earlier in the config

[rahul-tuli]