fix: harden local replica synchronization

[yunhaoli24]

Summary

• Route HTTP traffic from all projects through one process-wide queue, honor Retry-After globally after HTTP 429, and reuse recent history updates to avoid redundant file-tree diff requests.
• Debounce local watcher bursts and retry transient live uploads with bounded exponential delays.
• Prevent destructive startup fallback: full remote-to-local sync only runs for an empty replica or when local hashes still match the checkpoint.
• Pause synchronization without modifying either side when local and remote changes overlap, including parent-directory delete/rename conflicts.
• Ignore dot-prefixed paths such as .output and all symbolic links before filesystem processing.
• Fix the OT update direction for complete-file writes and report rejected Socket.IO updates with usable error details.
• Restore the local-replica creation flow when a remote project is open, and detect stale persisted replica directories.
• Rebuild local replica SCM state from .overleaf/settings.json after authentication expires, including reconnecting immediately after re-login.
• Stop the reconnect loop caused by disposing the provider-owned VFS after successful local replica creation.
• Isolate background project initialization from workspace-global collaboration, chat, history, SCM, status, and compile registrations.
• Keep deterministic workspace-feature failures out of the connection retry loop.
• Add structured connection and SCM-creation diagnostics with project ID, scheme, retry attempt, base URI, and original errors.
• Keep sync-state writes batched and in place so .overleaf/sync-state.json does not churn.

Verification

• npm run compile
• npm run lint (0 errors; 1 pre-existing warning in pdfViewEditorProvider.ts)
• git diff --check