Support multi-control edit datasets

[Xabsurd]

Summary

Adds support for multi-control edit datasets by allowing a dataset directory to specify multiple control image directories via control_paths.

This enables training edit models where each target sample is conditioned on more than one reference/control image, while preserving the existing single-control control_path behavior.

Changes

• Add control_paths = [...] support to directory dataset configs.
• Validate that control_path and control_paths are mutually exclusive.
• Store and cache multi-control tensors with shape [batch, num_controls, ...].
• Update Flux2 edit training to encode and consume multiple control images.
• Update Qwen-Image edit training and prompt processing to support multiple VLM/control images.
• Document multi-control support and add an example config snippet.

Example

[[directory]]
path = '/path/to/targets'
control_paths = [
  '/path/to/control_00',
  '/path/to/control_01',
]