Make `Processor.training_step` abstract.

[cisprague]

In contrast to Processor.validation_step – which should mostly be the same for all processors (compares rolled out trajectory to true trajectory) – Processor.training_step seems to depend more on the specific subclass of Processor. E.g., the flow matching loss can be different from the diffusion loss.

Related to this: I think we can get rid of

auto-cast/src/auto_cast/processors/base.py

Line 21 in 1a669a4

loss_func: nn.Module | None = None,

as, in most cases, the loss function will be specific to the subclass.