Snapshot cache

See merge request !2

This will eventually allow us to use --previous and skip the expensive
remote snapshot lookup.

Cache directories are sharded by source name, enabling re-use.

Avoids an expensive lookup of the previous snapshot.

Local caches apparently require an exclusive lock, effectively
preventing concurrency -- it is still to be seen if the repository
lock has the same effect though.

V2

See merge request !1

This way we can detect failure of dump / restore commands when used
with the 'pipe' handler (when we're using shell pipes for composition).

Clear separation between configuration and its parsed results, which
are now maintained in the RuntimeContext to kept them consistent
within backup/restore jobs.

This makes it easier to generate application-level atom paths that are
manageable. Datasets are given unique IDs at generation time.

Use LZ4 by default.

It may be set on the atom instead.