[Feat] Add support for SGLang quantization

Check model type and organize code.

Restore setup.py.

Backup weight_scale in cpu memory and restore them after reloading, because all weights will be set as zero by sglang if enable_memory_saver is set as True.

Clone existing params instead of creating empty tensors to fix the bug of qkv_weight caused by loading qkv respectively.

Clean python and torch cache only once per updating weights.

Only operate on updated params for batch weight updates.

Modify code and verify weights reloading of sglang.

Finish patched load_weights_and_postprocess, the program can run now.

Patch load_weights_and_postprocess of sglang; Bugs exist.

Patch sglang Engine.

Merge pull request #35 from sallyjunjun/patch-1

add load_column_parallel_weight to recorded_loader_keys

Clarify installation instructions in README

Add manual patch section to README