Handle gradient accumulations at the end of epoch differently about lightning

[awaelchli]

@jakub-dyno If the epoch size is not evenly divisible by the gradient accumulation size, the optimizer will step anyway on the last iterations. The partial accumulation can't be kept around in memory and continued in the next epoch.

The loss over the accumulation window is scaled regardless of whether it is the full window or not:

pytorch-lightning/src/lightning/pytorch/loops/optimization/automatic.py

Line 327 in e330da5

return self.output_result_cls.from_training_step_output(training_step_output, trainer.accumulate_grad_batches)

pytorch-lightning/src/lightning/pytorch/loops/optimization/automatic.py

Line 81 in e330da5

closure_loss = closure_loss / normalize

If you'd like to change it, a PR for this would be welcome. But it could be a bit involved.