fix(data-store): use REPEATABLE READ isolation for file metadata queries

[Theodus]

Prevents queries from missing rows due to concurrent table modifications
during compaction/GC operations.

1. Refactor streaming API to Vec (get_revision_files)
   • All callers were immediately collecting the stream into a Vec
   • Simplifies the API and makes transaction handling straightforward
2. Add REPEATABLE READ transaction isolation
   • Wrap file metadata query in a transaction with REPEATABLE READ
   • Guarantees consistent snapshot from transaction start

Concurrent compaction/GC operations could cause the file metadata query
to miss rows, leading to spurious "missing block" detections that
triggered redundant re-dumping.

[JohnSwan1503]

Looks good to me.