Refactor the quantization config for weights

[thincal]

Feature request

Currently the quantization info is hacked into the weights in many ways, result in not easily adding more features with quantization:

1. using a special function _set/get_gptq_params, but which is already applied for awq and eetq, this is very confusing for maintaining the code
2. to support multi quantize version for awq, this version info is missed in current implementation
3. to load a quantized weights for inference such eetq method, there is no elegant way to know the weights are whether quantized or not.

So that we could simply binding the config info with weights, and inside weights getting the desired info directly.

Motivation

A refactor for supporting quantization fetures.

Your contribution

Yes, I am prepared a PR.