Refactor hyper-parameters of adapters and model loader

[hiyouga]

• We deprecate the checkpoint_dir argument and use the adapter_name_or_path argument to load peft adapters and rename resume_lora_training to create_new_adapter with reversed values.
• We create llmtuner.model.patches module to disentangle the patches from the model loader.

[Katehuuh]

No reference in code for --adapter_model_or_path, so is --adapter_name_or_path.

• We deprecate the checkpoint_dir argument and use the adapter_model_or_path argument to load peft adapters and rename resume_lora_training to create_new_adapter with reversed values.

[hiyouga]

@Katehuuh thanks for pointing it out