fix: cancel pending request tokens after transport EOF

[alishobeiri-oai]

Pending request handlers can outlive transport EOF while a caller retains the RunningService or a Peer. Their RequestContext cancellation tokens are children of the service loop token, but the EOF path did not cancel that token.

Cancel the service loop token after the existing response-drain period and before closing the transport. This preserves graceful response draining and wakes remaining handlers without cancelling an externally supplied parent token or sibling services.

The regression sends two requests over a real in-memory transport, leaves both handlers pending, and closes the peer input. It retains the service and peer while asserting that both request tokens are cancelled and both handlers finish. Tokio time is paused so the drain deadline adds no wall-clock delay.

Validation

• The new regression fails without the service-token cancellation.
• cargo +1.97.1 test -p rmcp --test test_inflight_response_drain --test test_close_connection --test test_cancelled_response --features "client server transport-io": all 16 tests passed.
• Formatted the changed files with rustfmt.

[alexhancock]

Good catch