fix: stop DO changes from being lost when a peer misses a sync

[pjwerneck]

Summary

DO clients that share a SyftBox folder share one on-disk hash cache, and a sync recorded a change as seen before it sent the change. A peer that the version gate skipped, or whose upload failed, never got the change. This PR keeps each change pending per peer until the upload to that peer succeeds.

Changes

• Seen with syft-bg on Colab: apis and auto-approved job results never reached the submitter, and a later do.sync() sent nothing
• The DO records each queued path as pending per peer in .cache/owner_pending_deliveries.json, and clears it after the upload to that peer succeeds
• Approved peers that fail the version gate are deferred: their changes stay pending and go out in the first sync where the peer passes
• Each sync resends the current state of pending paths (a deletion if the file is gone) and drops paths the peer can no longer read
• A pending file that exists but cannot be read raises instead of going out as a deletion
• PersistedDict.update_entry: read-modify-write under one exclusive lock

Testing

• tests/unit/test_persisted_cache.py: a second DO process (syft-bg) syncs first and skips the peer, fails the upload, defers a deletion, or hits a read error; the peer still ends up with the right state. All four fail without the fix

Asana task