Dev gan trainer

[wli51]

This PR adds back the Generative Adversarial Network (GAN) training functionality following its removal in refactor #19 where the model forward pass logic is decoupled from trainer and promoted as a separate engine subpackage (but #19 only included the engine parts for regular UNet training).

Introduces

1. src/virtual_stain_flow/engine/forward_groups.py:
   Added DiscriminatorForwardGroup, a standardized interface for discriminator forward passes

2. src/virtual_stain_flow/engine/orchestrators.py:
   An additional layer of abstraction. OrchestratedStep and GANOrchestrator between the forward passing engine and the trainer classes to define complex orchestrations involving more than one model (neural network), as needed when training GANs.
   Specifically, GANOrchestrator provides training interface to separately train the generator and discriminator

• with complexity coming from the fact that when training only the generator, the data also has to flow through the discriminator, and vice versa
• and that the generators needed to be updated less frequently than the discriminator to ensure stable training
  hence the orchestrator abstract helps makes things cleaner.

3. New operations for Context class

4. Minimal tests

[d33bs]

I double checked and found the tests all pass, nice job. Given the great stuff going on with testing you might consider making testing a part of automated jobs with a GitHub Actions job. The benefit here is that it'd provide you and reviewers confidence about the changes within the context of the pull request (similar to how the pre-commit checks operate).