This PR implements ChromaTransformer2DModel and ChromaPipeline as separate classes, addressing the feedback from DN6 about creating dedicated implementations rather than using variants within FluxTransformer2DModel.

Changes Made:

Core Implementation

• ChromaTransformer2DModel: New transformer class specifically for Chroma, separate from FluxTransformer2DModel
• ChromaPipeline: Dedicated pipeline without CLIP encoder (T5-only) for Chroma's requirements
• Single file converter: Added support for loading Chroma checkpoints from safetensors files
• Attention masking: Implemented proper attention masking for Chroma's specific requirements

Package Integration

• Updated all __init__.py files to export new classes
• Added proper imports in src/diffusers/__init__.py
• Integrated with existing single file loading infrastructure

Example Usage

• examples/chroma_generation.py: Comprehensive example script showing how to use Chroma model
• Demonstrates both HuggingFace Hub and single file loading methods
• Includes optimization techniques (CPU offloading, VAE slicing/tiling)

Technical Details:

• Uses T5 XXL encoder instead of CLIP
• Compatible with FLUX VAE
• Supports FlowMatchEulerDiscreteScheduler
• Implements proper device and dtype handling
• Memory-efficient optimizations included

This implementation provides a clean, separate architecture for Chroma while maintaining compatibility with the existing diffusers ecosystem.