Standardize usage of `batch_size`

[lballes]

Currently, our ER-based methods use a batch that consists of batch_size points from the current task and memory_batch_size points from the memory. This is inconvenient to compare/standardize to some other learners (e.g., Joint, GDumb, Fine-tuning) and will also result in a smaller total batch size in the first training stage (when no memory exists).

I propose that all methods use batches of size batch_size. ER-based methods can have an additional argument called memory_frac that determines which fraction of the batch will be filled with points from the memory.