[Rust][Arrow] Preserve ACK deadlines during graceful rotation

[teodordelibasic-db]

Summary

Server-requested Arrow Flight graceful rotation currently replaces normal ACK waiting with the server grace-period deadline. A pending batch can therefore outlive server_lack_of_ack_timeout_ms while graceful close or transport drain is active.

Keep the oldest pending batch's absolute ACK deadline active throughout graceful rotation.

Prior draft #592 explored this behavior on an obsolete effort/* stack. It was closed so the implementation can be rebuilt as a fresh PR stacked on #654.

Scope

• Select among the next response, the oldest pending ACK deadline, and the graceful-close deadline while waiting for pre-rotation ACKs.
• Keep the ACK deadline armed during bounded request/response transport drain.
• Preserve response-first exact-tie handling while allowing only one ready response to defer an already-expired pending head.
• Re-check the same head under lock before returning DeadlineExceeded.
• Document that graceful rotation does not pause or extend server_lack_of_ack_timeout_ms.

Acceptance criteria

• An expired pending batch followed by a zero-progress graceful-close signal fails with DeadlineExceeded.
• A batch whose ACK deadline expires during transport drain also fails with DeadlineExceeded.
• An ACK already ready at the deadline can complete the relevant pending head.
• Multiple ready responses cannot repeatedly defer one expired head.
• Existing graceful-rotation and H2 half-close tests remain green.

Non-goals

• Pending-relative deadline calculation itself, which belongs in prerequisite [Rust][Arrow] Make ACK deadlines pending-relative #655.
• General H2 request half-close and drain mechanics, which were addressed by [Rust] Fix Arrow Flight rotation H2 resets #649.
• Explicit client close() during recovery; tracked by [Rust][Arrow] Coordinate explicit close with recovery #657.