Feat #12: Add support for dynamic model loading with --model flag to test other VLMs

[aditi-dsi]

This PR enables dynamic loading of different VLMs based on the --model argument.

• Gemma models use the native Gemma3ForConditionalGeneration class.
• Non-Gemma VLMs fallback to appropriate Hugging Face model classes dynamically.
• We can add more models & model_classes as required to model_class_map to extend support for testing other VLMs.

Additionally, added a minor fix in train.py:

• fixed project_name to fallback to None when no project name is given (which otherwise interrupts the training with an error).

[aditi-dsi]

@sergiopaniego this PR is ready for review.

[ariG23498]

I think AutoModelForImageTextToText should entail everything. Could you give it a try?

[aditi-dsi]

Sure, taking a look, will let you know.