[Config] Investigate config refactor for cleaner serialization, reduced logic

[kylesayrs]

Background

Right now, quantization configs are serialized through the following lifecycle:

1. apply_quantization_config is used to attach quantization_scheme attributes to modules
2. The model undergoes calibration and compression
3. The quantization config is regenerated from the model using QuantizationConifig.from_pretrained
4. The new config is serialized by ModelCompressor.update_config

This approach has some downsides (see phi3 example config)

• Any config group names set by the user are discarded
• The config groups which are generated do not necessarily match the config groups set by the user
• The ignore list becomes very large and ugly to read
• The logic for generating a config from a model is very difficult to maintain

The scope of this issue to investigate an approach whereby step (1) attaches the config as a quantization_config attribute on the model, which is then read by step (4) without having to go through step (3). This would mitigate all of the above downsides.

! Some things to keep in mind

• apply_quantization_config may be applied multiple times. This may necessitate some logic to "merge" quantization configs. This has been written as a draft already, feel free to ping @kylesayrs if you would like to leverage this, or feel free to use your own.

[dsikka]

@kylesayrs is this still relevant?