fix: clean up leaked observers, untracked tasks, and misleading log

[datlechin]

Summary

Follow-up to #217. Full resource cleanup audit found 3 additional issues:

• URL filter notification observers never removed — setupURLNotificationObservers() registered two block-based NotificationCenter observers but never stored the returned NSObjectProtocol tokens. Each window open/close accumulated 2 orphaned observers. Now stores tokens and removes them in teardown().
• Untracked metadata driver creation Tasks — connectToSession() and reconnectSession() spawned fire-and-forget Task {} blocks to create metadata drivers that were never stored or cancelled. If disconnectSession() ran before completion, the task could assign a driver to a stale session. Now tracked in metadataCreationTasks and cancelled on disconnect.
• Misleading health monitor log message — Log said "after 3 retries" but ConnectionHealthMonitor retries indefinitely with exponential backoff. Fixed the message.

Also adds docs/development/resource-cleanup-audit.md tracking all findings from the audit.

Closes #214

Test plan

• Build succeeds
• Connect via SSH tunnel, close window, verify no leaked processes
• Open/close multiple connection windows, verify no observer accumulation
• Disconnect while metadata driver is still loading, verify no stale assignment

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.