Allow GPTQModel to auto select Marlin or faster kernels for inference only ops

[LRL-ModelCloud]

No description provided.

[Qubitium]

This PR addresses the issue where GPTQModel is unable to auto-select the fastest quant linear (Marlin) due to missing information regarding if this model is loaded via pre-trained or from_quantized. Add pack=True requirement to hf_select_quant_linear (for pretrained quantization will solve this. Without the PR, user need to manually set backend="marlin" to use Marlin. This PR allows auto selection of Marlin, if compatible, in default auto mode.

[Qubitium]

@SunMarc @IlyasMoutawwakil Ready for review. Unlike AutoGPTQ, GPTQModel has pretty robust internal code to select the best/fastet quant linear available for model inference. However, for this to happen, we need to receive signal if pack: bool (packing) will be used. For loading a post-quantized model, packing is not needed so we can auto select Marlin on Nivida CUDA available platform or next-best kernels for the current env. Some fast inference kernels are inference only and cannot perform packing.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[SunMarc]

LGTM !