(#2510) allow mask conditioning_type to work on edit models that require latent conditioning

[bghira]

Closes #2510

This pull request refactors how conditioning datasets are handled and routed during training, especially for edit models and loss masking. The main improvements are a clearer separation of reference and mask/segmentation conditioning types, more robust validation for Qwen edit models, and consistent use of a new loss_mask_type field throughout the codebase. This enhances flexibility and correctness when working with multiple conditioning datasets.

Key changes include:

Conditioning Dataset Handling & Routing:

• Refactored the collate logic in simpletuner/helpers/training/collate.py to distinguish between reference conditioning types (reference_strict, reference_loose) and mask/segmentation types (mask, segmentation). This allows separate routing for model input (reference) and loss masking (mask/segmentation), supporting more complex dataset setups. [1] [2] [3] [4] [5]
• Added new fields conditioning_type and loss_mask_type to the prepared batch output, making the distinction explicit for downstream code. [1] [2]

Qwen Edit Model Validation:

• Updated Qwen edit model validation to require at least one reference conditioning dataset, instead of enforcing that all conditioning datasets be of reference type. This allows additional mask/segmentation datasets for loss masking, improving flexibility and correctness. [1] [2]

Loss Masking Logic:

• Standardized all model and loss functions to use loss_mask_type instead of conditioning_type when applying mask/segmentation-based loss masking, ensuring consistent behavior across models. [1] [2] [3] [4] [5] [6]