[Fixing] Better vision model finetuning

[danielhanchen]

• [ ] Allow mixing text and vision data (ie rows of data without any images
• [ ] Resizing images automatically by checking the preprocessor, since memory usage can explode on large images. See https://github.com/unslothai/unsloth/issues/1524#issuecomment-2584971126
• [ ] Allow saving to GGUF for Llava type models
• [ ] Unsure exporting to 16bit does not miss any files for eg https://github.com/unslothai/unsloth/issues/1521
• [ ] train_on_responses_only for VLMs