EXTRACT: Efficient Policy Learning by Extracting Transferrable Robot Skills from Offline Data