Introduce new parameter to control number of batches in negative edge sampling

[einbandi]

The parameters n_epochs and batch_size cannot fully define the number of batches in negative edge sampling. Since the negative edge sampling uses a weighted random sampler without replacement, it cannot be guaranteed that each item is seen exactly once. Instead, the user must define something like a batches_per_epoch parameter (rethink name, maybe check how it is called in UMAP).

[einbandi]

This is actually not necessary, it is enough to just set a higher number of epochs. It is now explained in the docstring of the TrainingPhase class how the batch size relates to the number of edges and items in a batch. There, also the different meaning of epoch in the case of the two sampling variants is explained.

[einbandi]

On second thought, it might be better to have a batches_per_epoch parameter that by default is set in such a way that roughly a number of items/edges is samples per epoch is sampled that is equal to the full dataset (but can be overridden).