earlyMessageRejectionPeriod can drop messages unexpectedly when tasks run long

To ensure that replica Kafka tasks are making decisions in sync with each other, earlyMessageRejectionPeriod works by setting a maximumMessageTime on the Kafka tasks spawned by a supervisor, based on the configured taskDuration of the task. However, in some situations, tasks can run longer than their configured duration:

- The remainingSeconds counter resets when supervisors fail over
- Supervisors might not be running when the task duration arrives
- Replica tasks may continue running for some time after the task duration arrives, due to the checkpoint-and-sync protocol

In these situations, if earlyMessageRejectionPeriod is too low, messages will be dropped unexpectedly.

I'm not sure what the best way is to fix this -- we do want a fixed maximumMessageTime in order to ensure that all replicas are in sync.

Start by tracing how earlyMessageRejectionPeriod sets maximumMessageTime for Kafka tasks spawned by a supervisor, then examine the remainingSeconds reset, supervisor downtime, and checkpoint-and-sync cases described in the issue. Define a solution that preserves replica synchronization without unexpectedly dropping messages when tasks outlive their configured duration.

Written by the indexing model from the issue text.