LLM: 2bit quantization support

[rnwang04]

Description

• should merge after https://github.com/intel-analytics/llm.cpp/pull/232

0. Workflow

• add support for importance matrix and related preprocess
• add iq2_xxs/ iq2_xs support in load_in_low_bit parameter and related quantization workflow update
• verification of iq2_xxs / iq2_xs quantization
• add dequantization support for iq2_xxs
• verification of iq2_xxs dequantization support
• rest token inference with iq2_xxs

1. Why the change?

To support 2bit quantization.
https://github.com/analytics-zoo/nano/issues/1015
This PR is just a starting point of iq2 support, which just add basic function support, will take a look into q2_K mixed quantization later.

2. User API changes

Related imatrix is downloaded from https://huggingface.co/datasets/ikawrakow/imatrix-from-wiki-train/tree/main .

from bigdl.llm.transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained('meta/Llama-2-7b-hf',
                                             load_in_low_bit='iq2_xxs',
                                             optimize_model=False,
                                             trust_remote_code=True,
                                             use_cache=True,
                                             imatrix='imatrix/llama-v2-7b.imatrix').to('xpu')
model = AutoModelForCausalLM.from_pretrained('mistral/Mistral-7B-Instruct-v0.2',
                                             load_in_low_bit='iq2_xxs',
                                             optimize_model=False,
                                             trust_remote_code=True,
                                             use_cache=True,
                                             imatrix='mistral-7b-instruct-v0.2.imatrix').to('xpu')

3. Summary of the change

• Support loading and preprocess importance matrix which is produced by llama.cpp's format
• Provide new ggml_quantize_tensor_with_weights API
• Support basic functionality for iq2_xxs mixed with q4_0

4. How to test?

• Unit test
• Local test