JMX metrics stuck when connector graceful exit fails

When connector graceful exit fails, the old JMX metrics get reported even after the same
connector is restarted and successfully producing events.

Example scenario:

• Connector is starting up, reading history topic from kafka.
• During the startup configuration is updated or some exception occurs at startup time.
• Proper exit fails with logs like this: "Graceful stop of task job_1028-0 failed. [org.apache.kafka.connect.runtime.Worker]"
• Now when task or connector is restarted the JMX metrics are stuck with old values, which makes monitoring health very difficult.