fix(merge_model): use torch_dtype for MLCDVisionModel and remove invalid device_map from CLIPImageProcessor

[inigopm]

Problem

Two bugs in validate_vit_consistency() introduced by #52 and only partially addressed in #88:

1. MLCDVisionModel.from_pretrained — missing torch_dtype=torch.float32. The original code used torch_dtype= correctly; refactor(merge_model): Enhanced file implementation robustness #52 changed it to dtype= (silently ignored by from_pretrained); Update merge_model.py #88 removed the param entirely, losing the explicit float32 cast.

2. CLIPImageProcessor.from_pretrained — still has device_map=... after Update merge_model.py #88. CLIPImageProcessor is a ProcessorMixin, not a PreTrainedModel; device_map is not a valid argument and raises TypeError at runtime.

Fix

# before
rice_model = MLCDVisionModel.from_pretrained(vit_path, device_map={"": f"cuda:{CUDA_DEVICE}"})
processor  = CLIPImageProcessor.from_pretrained(vit_path, device_map={"": f"cuda:{CUDA_DEVICE}"}, use_fast=True)

# after
rice_model = MLCDVisionModel.from_pretrained(vit_path, device_map={"": f"cuda:{CUDA_DEVICE}"}, torch_dtype=torch.float32)
processor  = CLIPImageProcessor.from_pretrained(vit_path, use_fast=True)