BtS accumulation with non-multiples

[MaxiBoether]

Currently, Batch-Then-Sample is limited to batch sizes and selection ratios where batch size % batch_size * ratio == 0, e.g. 50% selection ratio. This is because from each batch we select ratio * size samples, and then build a new batch after n batches.

In Sample-Then-Batch, we don't have this limitation. It's a bit unclear to me whether we can find a solution in which the number of samples we train on is exactly identical between StB and BtS. One idea for BtS is to round the target size per batch (up or down?), and then at some point we selected a little too much points. We can use them for the next batch. However, it is not clear to me if this will lead to identical behavior to StB. We should synchronize the number of trained batches to be able to fairly compare BtS and StB.