Process threads

[Qubitium]

Refractor multi-gpu quantization with python threads. Compare to previous efforts:

• 2-3x less memory usage if you assign multiple GPUS up to 3 for Llama/Qwen style or 5 (tested) for MoE style.
• Performance hit is much reduced verious pervious method. Small model only saw 1% slowdown.
• This is not tensor-parallel so it will not make things faster, but actually around 1-5% slower dependng on model, gpu, and gpu to gpu memory latency.

With advent of python 3.13-3.14 and pytorch 2.7 concurrent gil context, it is possible to get Nx GPU improvement to MoE quantiation for next major refractor if we can cleanly decouple the thread/module code with multiple python/torch GIL context without any usage of tensor-parallel which I want to avoid at all-costs for now.