Add HQQ model serialization support

[mobicham]

What does this PR do?

The goal of this PR is to add support for saving/loading transformers models quantized via HQQ.
Currently, it's not possible to directly save/load HQQ-quantized models via model.save_pretrained. The workaround is to use the hqq lib which offers serialization but without safetensors support.

The first thing to do was making the HQQLinear.state_dict compatible with safetensors: dropbox/hqq@74bbe01
Since safetensors only support torch.Tensor dtype and no nested dictionaries are allowed, I had to implement an encoding/decoding logic for the parameters, which works fine for the moment.

The second step was integrating the new HQQLinear state_dict loading with the transformers lib, which I found a bit tricky because of the following issues:

• Some state_dict parameters are split across 2 safetensors files. So loading fails for that specific layer since I can only get the parameters from the first file. For now, I put a fake quantized layers so that loading doesn't break but you'll see a message.
• dispatch_model breaks with HQQLinear module, I am not sure what's the issue here. Before it reaches that function, the layer is already loaded from the state_dict. If I skip the dispatching function, I a get a model with meta device for all the layers except HQQLinear

Since this is a draft pull-request, I am commenting out some stuff while debugging, will fix that later (you'll see there's a TODO @mobicham comment so I can get back to it).

Will need to drop support for quantized scale/zero, so I will need to update the HqqConfig and the documentation, as well as add tests later.

Any help is highly appreciated , thank you!

Who can review?

@SunMarc

Example

import torch, gc
from transformers import AutoModelForCausalLM, AutoTokenizer, HqqConfig

quant_config  = HqqConfig(nbits=4, group_size=64, quant_zero=False, quant_scale=False, axis=1)

model = AutoModelForCausalLM.from_pretrained(
    'meta-llama/Meta-Llama-3-8B', 
    torch_dtype=torch.float16, 
    cache_dir='/nas/hicham/tmp/',
    device_map="cuda:0", 
    quantization_config=quant_config
)
 
#Test 
with torch.no_grad():
	out = model.forward(torch.zeros((1, 8), dtype=torch.int32, device='cuda:0'))


# Save
model.save_pretrained("llama3-hqq")
del model
torch.cuda.empty_cache(); gc.collect();

#Load
model_loaded = AutoModelForCausalLM.from_pretrained(
    'llama3-hqq', 
    torch_dtype=torch.float16, 
    cache_dir='.',
    device_map="cuda:0")

## Does't work here because of `dispatch_model()`

with torch.no_grad():
	out = model_loaded.forward(torch.zeros((1, 8), dtype=torch.int32, device='cuda:0'))

[amyeroberts]

Gentle ping @SunMarc