Force equal batch sizes for `ReshufflingBatchSubsampling`

[Red-Portal]

Turns out DifferentiationInterface.prepare_gradient requires the batch sizes to be exactly equal. Currently, if the number of samples does not exactly divide by the batch size, the last batch might end up having a different size. This needs to be fixed if we want to use prepare_gradient in a subsampling context.

[gdalle]

I agree that it's not ideal but some backends have size-specific preparation mechanisms (e.g. ForwardDiff). An alternative would be to prepare the last batch separately