S3 collab cache

[Horusiath]

This PR reimplements DiskCache to store encoded collabs in S3-compatible storage rather than in Postgres. Postgres is used only as a fallback for compatibility with an existing storage model.

TODOs:

• Use S3 for collab snapshot data as well.
• We need to decide if we really want to store EncodedCollab format in S3. IMO it's useless - it's heavily reliant on bincode (which is not version tolerant) and all of the fields (namely: state_vector) can be resolved from doc_state directly instead. We could keep a straight up yrs Doc state instead.
• Added ZSTD compression of S3 objects.
• Added a threshold (S3_COLLAB_THRESHOLD, 2000 bytes), so that any collab data smaller than that threshold will still land in Postgres. PS: All new snapshots are always stored in S3 regardless of size.