Fix GPU memory allocation issues

[Ednaordinary]

Previously, the following would not properly clear all memory:

import gc, torch
del pipeline, transformer
torch.cuda.empty_cache()

This PR fixes that by resetting the internal quantized model state on __del__ (please note this is not actually called until the ref count for the transformer is 0, but this matches the behaviour of other projects like diffusers)

Since this calls the full reset() method of ModuleWrapper, it may have unintended consequences if multiple models are loaded at once (it's probably fine though)

[Ednaordinary]

This may solve the following:

nunchaku-tech/ComfyUI-nunchaku#65 should be fixed by this PR (I do not use comfy ui, dunno how it manages models)

I believe nunchaku-tech/ComfyUI-nunchaku#57 is caused by memory pinning, a separate issue, which was an issue I had where pinned memory could not be allocated but still took up memory (even after the process had exited, had to fully reboot).

Can be solved by telling nunchaku not to pin memory:

import os
os.environ['NUNCHAKU_LOAD_METHOD'] = 'READNOPIN'

I am less sure about comfyui specifically since I don't use it, but providing the variable on server start may work

export NUNCHAKU_LOAD_METHOD='READNOPIN'
python3 main.py

[lmxyy]

Thanks for the PR. Looks good to me @sxtyzhangzk Could you take a look?

[lmxyy]

Passing all tests, approved.