[Enhancement] Add flexible projector selection (Linear, MLP, Attention) for multimodal alignment

[nguyencongtuyenlp]

👋 Hi @haotian-liu, this PR is ready for review.

Adds flexible projector options (Linear, 2-layer MLP, lightweight Attention).

Keeps backward compatibility with linear.

Verified initialization and training for all projector types without breaking existing configs.

Would appreciate your feedback when you have time 🙏

This PR introduces a new projectors.py module and updates llava_arch.py to allow flexible selection of projector types for image-text alignment.

• Added support for Linear, 2-layer MLP, and lightweight Attention-based projectors.
• Updated LlavaMetaModel to use the new build_projector() API.
• Kept backward compatibility (linear as default).

Motivation:
Different projector architectures may improve performance depending on dataset size and modality alignment. This change allows the community to experiment more easily.

Tested:

• Verified training initializes with each projector type.
• Code runs without breaking existing configs.

Fixes #1884