fix(server): propagate errors through TimeoutMiddleware

[hsusul]

What does this PR address?

When an HTTP handler raised before sending any ASGI response, TimeoutMiddleware waited for the full traffic.timeout and then returned no response. Clients received an empty disconnect instead of a normal 500 response.

The timeout waiter previously completed only when:

• the application sent its first ASGI message, or
• the timeout expired.

If the application future had already failed, fut.cancel() returned False. The middleware then neither sent a 504 response nor awaited the future, leaving the exception unobserved.

This PR:

• completes the waiter when the request future finishes before sending a response
• awaits the request future when cancellation loses the timeout race, allowing the exception to propagate through Starlette's ServerErrorMiddleware

Successful responses and genuine timeout responses are unchanged.

Validation

• Added regression coverage for an exception raised before response headers are sent
• Verified slow requests still return 504
• Verified successful requests still return 200
• ruff format passed
• ruff check passed
• pre-commit passed on changed files
• pyright completed with 0 errors
• git diff --check passed

export PYTEST_PLUGINS=bentoml.testing.pytest.plugin
python -m pytest tests/unit/_internal/server/http/test_traffic.py -q