Validation timesteps

[dxqb]

Implementing the conclusions of this thread: https://github.com/Nerogar/OneTrainer/issues/772

Summarized:

• Validation on timestep 500 is not ideal, but hardcoded currently
• Choosing validation timesteps from a distribution is not good either, especially for small validation sets
• [X] let the user choose
• There is no meaningful way to average the loss of different timesteps
• [X] report separately to tensorboard

I find it helpful to validate on a high timestep that determines image composition, and something in between - but other best practices might evolve