support multiple storage devices, including memory-only storage for use in unit tests

[haydnv]

When this issue is resolved it should be possible for a library to:

1. automatically move data from host memory to a specific accelerator device memory (e.g. on a host with two GPUs, data not in active use on GPU Update repository link #1 should be bumped back to main memory and all data in active use on GPU allow access by a key type which implements PartialEq<str> + PartialOrd<str> #2 should reside on GPU allow access by a key type which implements PartialEq<str> + PartialOrd<str> #2)
2. automatically minimize data movement between all devices
3. implement unit tests which do not require filesystem access even for code which depends on freqfs filesystem access methods