FEAT: Add possibility of skipping modules when quantizing

[younesbelkada]

What does this PR do?

For some models (e.g., Whisper, Mixtral or Llava) it is important to skip some modules during quantization. This PR adds the experimental support for skipping modules during quantization with the following API.

from awq import AutoAWQForCausalLM
from transformers import AutoTokenizer, AwqConfig

model_path = "facebook/opt-125m"
quant_path = "test-quant/opt-125m-awq-no-kproj"
modules_to_not_convert = ["k_proj"]

quant_config = {"zero_point": True, "q_group_size": 128, "w_bit": 4, "version":"GEMM", "modules_to_not_convert": modules_to_not_convert}

# Load model
model = AutoAWQForCausalLM.from_pretrained(model_path)
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)

# Quantize
model.quantize(tokenizer, quant_config=quant_config, modules_to_not_convert=modules_to_not_convert)

An example model has been pushed here: https://huggingface.co/ybelkada/opt-125m-awq-no-k-proj and works fine with a PR of transformers that I will open soon.

cc @casper-hansen @TheBloke