Solve KeyError: 'pixel_values' in Qwen-Image-Edit-2509 when number of control images is less than 3

[nicolaus-huang]

fixes #441: Standardize control image handling for training pipeline

This commit refactors the way control images are processed and passed to the model to ensure consistency and robustness.

• In extensions_built_in/diffusion_models/qwen_image/qwen_image_edit_plus.py:

  • A new method, _prepare_control_images_for_prompt, has been added at line 158.
  • This function unifies control images from the data loader—whether a single tensor or a nested list—into a standard list format ([tensor_ctrl0, ...]).
  • It is called within get_prompt_embeds, where each image is then interpolated, has its extra batch dimension squeezed, and is passed to the Qwen pipeline as a CPU float32 tensor.

• In extensions_built_in/sd_trainer/SDTrainer.py:

  • The prompt encoding arguments at lines 1091 and 1484 have been adjusted.
  • The training process will now prioritize using batch.control_tensor_list.
  • If it is unavailable, the code will fall back to splitting batch.control_tensor into a list. This ensures control images are always passed to the model during training.