[rpc] Missing client-side request deadlines can stall Flink writes until connection idle timeout

### Problem

Java client RPCs have no independent timeout while pending or in flight. If a request receives neither a response nor a connection failure, its future can remain incomplete until the connection closes, delaying retries and blocking Flink sink flush/checkpoint completion during a rolling upgrade.

For writes, `client.request-timeout` (default 30 seconds) is sent to the server for its acknowledgement wait; it does not enforce a client-side response deadline. Recovery may therefore wait for the default 10-minute connection idle timeout, or longer if other traffic keeps the connection active.

### Example reproduction

A transport-level reproduction, without requiring a rolling upgrade:

1. Start a localhost `NettyClient` and `NettyServer`. Set `client.request-timeout=100 ms`, client `netty.connection.max-idle-time=2 s`, and server idle timeout to 30 seconds.
2. Allow version negotiation to finish so the connection becomes ready.
3. Send an `API_VERSIONS` request whose test-server handler returns a deliberately uncompleted future. Keep the connection open.
4. After 500 ms, the client future is still incomplete and the request remains in `inflightRequests`, despite the configured 100 ms request timeout.
5. At approximately 2 seconds, the client closes the idle connection. Only then does the request fail with `NetworkException` and leave the in-flight map.

Expected: the RPC should expire through a client-side deadline and release its state independently of connection idleness.

Start with the Java NettyClient request path and its inflightRequests handling, using the API_VERSIONS reproduction described in the issue. Verify the configured 100 ms client request timeout against the 2-second connection idle timeout. Done means an unanswered RPC fails at its client-side deadline and is removed from in-flight state without waiting for connection closure.

Written by the indexing model from the issue text.