fix: release lock before streaming and add kernel interrupt support

[devin-ai-integration]

Summary

Fixes #213 — asyncio.Lock in messaging.py not released on client disconnect, causing cascading timeouts.

Two changes:

1. Lock scope narrowing (template/server/messaging.py): Splits execute() into Phase A (under lock: prepare env vars, send request, schedule cleanup task) and Phase B (no lock: stream results). When a client disconnects mid-stream, the lock is already released — no orphaned lock, no cascade. Addresses the env-var cleanup race flagged in PRs fix: release lock before streaming to prevent orphan on client disconnect #234/fix: release asyncio.Lock before streaming to prevent orphan on client disconnect #235 by creating _cleanup_task while still holding the lock.

2. Kernel interrupt endpoint (template/server/main.py, Python SDK, JS SDK): Adds POST /contexts/{id}/interrupt that proxies to Jupyter's kernel interrupt API. Exposed as interrupt_code_context() / interruptCodeContext() in SDKs. Allows stopping long-running code without restarting the kernel (state preserved).

Review & Testing Checklist for Human

• Verify concurrent websocket sends are safe with websockets==12.0: After the lock is released, _cleanup_env_vars may call self._ws.send() concurrently with the next execute()'s self._ws.send() (under lock). The assumption is websockets 12.x handles this internally — confirm this or add explicit serialization.
• Verify cleanup task doesn't re-introduce blocking: The cleanup task is await-ed at the start of the next execute(). If the kernel is still busy with the previous code, this await blocks. Confirm this is acceptable behavior vs. the old cascading-lock behavior.
• Test the lock fix end-to-end on a live sandbox: Run the reproduction from the issue — sleep(30) with a 5s SDK timeout, then immediately run print('hello') on the same context. The second call should not be blocked by an orphaned lock.
• Test interrupt preserves kernel state: After calling interrupt_code_context on a long-running execution, verify variables/imports from prior cells survive.
• Confirm async Python SDK auth header consistency: The async interrupt_code_context doesn't send X-Access-Token — this matches the existing async restart_code_context pattern, but differs from the sync version which does send it. Pre-existing inconsistency, but worth noting.

Notes

• This supersedes closed PRs fix: release lock before streaming to prevent orphan on client disconnect #234 and fix: release asyncio.Lock before streaming to prevent orphan on client disconnect #235 (same branch, fix/lock-orphan-on-disconnect). The key difference is scheduling _cleanup_task under the lock to prevent the env-var isolation regression that Codex flagged.
• The finally block in Phase B uses self._executions.pop(message_id, None) for defensive cleanup if the generator is abandoned by Starlette on client disconnect.

Link to Devin session: https://app.devin.ai/sessions/d709ebe9b3e14cea89be89c9c2faa29e
Requested by: @mlejva

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment and CI monitoring