PRISE: Learning Temporal Action Abstractions as a Sequence Compression Problem