Clarify in documentation how `epoch_step` is considered with splits

[jzombie]

This needs double-checking but I believe when sampling across different splits the epoch_step still increments forward (not starting from 0 when moving from train to val), which isn't necessarily problematic (but maybe it is) and it needs documenting.

[jzombie]

Per-epoch reproducibility should always use checkpoints, regardless, as that's the only way to guarantee all internal sampler states are lined up.