fix(extension): free WS message-callback handlers when requests finish

[pjdoland]

Summary

`WebsocketCopilotHandler._messageCallbackHandlers` was a dict keyed by request `messageId`, populated when chat / generate-code / inline-completion requests started but never pruned. Every long chat session leaked one response emitter + cancel token per turn for the lifetime of the websocket connection. Entries still in flight at disconnect pinned their state until the worker happened to return on its own.

This was flagged in the security audit because the leaked entries hold a reference to the response emitter, which retains chat-history content; in a long session that's PII / prompt content sitting in memory after it should have been collectable.

Solution

Worker threads now run through a new `_run_request_thread` wrapper that pops the messageId from `_messageCallbackHandlers` in a `finally` so the cleanup fires whether the coroutine returns or raises. All three thread-spawn sites (chat, generate-code, inline-completion) route through it. `on_close` also clears the dict so a long-running request left in flight at disconnect releases its emitter and cancel token immediately rather than at coroutine completion.

Testing

Six new pytest cases in `tests/test_ws_callback_handler_leak.py`:

• normal-return cleanup
• raise-and-still-cleanup
• unknown-id-is-safe (idempotent pop)
• multiple concurrent requests each clean up only their own entry
• `on_close` drops all in-flight entries
• `on_close` is idempotent

Verified regression detection by reverting `extension.py` to `upstream/main` and re-running: 5 of 6 tests fail.

`pytest tests/` (738 passing, +6 new), `jlpm tsc --noEmit`, `jlpm lint:check`, `jlpm jest` (154 passing). All four green.