At the moment, creating a new DRef requires the associated data to be in memory. This makes it annoying to persist datasets and be able to easily re-load them in a future session. We should provide a means to generate a new DRef from a file stored on a device, and provide some high-level API (maybe another kwarg to poolset) to expose this.