(#2470) multi-aspect input conditioning for kontext, flux2 and qwen edit

[bghira]

Closes #2470

This pull request addresses issues with handling conditioning data in the batch preparation process, particularly when dealing with nested lists that can arise from varying image aspect ratios. It also makes minor adjustments to training example handling and test mocks for improved robustness.

Batch conditioning robustness:

• Added logic to flatten nested lists in the prepare_batch_conditions method of model.py, ensuring consistent handling of conditioning data even when input images have different aspect ratios.

Training example handling:

• Modified the _pop_training_example_for_path function in collate.py to pass an explicit is_conditioning flag when the conditioning type is "reference_loose", improving clarity and correctness in downstream processing.

Test improvements:

• Updated the mock for check_latent_shapes in test_collate.py to accept arbitrary keyword arguments, preventing potential errors during testing.