feat: authenticate interpreter requests through envd

[devin-ai-integration]

Summary

Authenticate every protected interpreter request through envd GET /auth, replacing closed #342 with a fresh implementation on main. Reuse the HTTP client connection pool, but do not cache validation results: token changes and envd failures are checked on the next request. Health and CORS preflight remain public, E2B_LOCAL=true is the explicit development bypass, and merged Jupyter isolation is unchanged.

Requires https://github.com/e2b-dev/belt/pull/4374 and e2b-dev/E2B#1955 deployed before the template. Breaking: existing SDK installations must upgrade to the header-forwarding release before using the new template; publishing the SDK alone does not prevent older clients' context operations from returning 401. Tokenless sandboxes are rejected. Health checks alone do not verify envd supports authentication. Related: SDK-363.

Verification: 19 server tests pass on Python 3.10 and 3.12 (including real Jupyter transport); lint, formatting, and compilation pass. Nine updated regression cases fail with the old cache, proving repeated/concurrent requests revalidate and token changes/upstream failures take effect without a cache delay. Removing middleware makes the unauthenticated-context regression fail (200 instead of 401). Actual Firecracker lifecycle testing and latency measurements remain outstanding.

Link to Devin session: https://app.devin.ai/sessions/23861bced08343e59c14adf841498b12
Open in Devin Desktop: https://app.devin.ai/desktop/session/23861bced08343e59c14adf841498b12?variant=devin
Requested by: @mishushakov

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

• Disable automatic comment, CI, and merge conflict monitoring