Improve pulsar source stability when using partitioned topics

Currently, Pulsar has a default failover timeout of 1s, when combined with partitioned topics in order to allow for distributed indexing, it is possible for indexers to become unstable and unable to make progress in the event that there is high network latency or an external factor causing the failover to trigger.

This is because we rely on the checkpoints to ensure the exactly-once semantics, which currently will error and restart the whole pipeline when topic partitions are assigned to indexers (this happens whenever an indexer connects/disconnects), when this restart occurs if the pipeline does not restart within 1s or so, Pulsar will re-assign the partitions and replay unacknowledged (documents not yet published by quickwit) causing a cycle to begin of partitions being assigned, re-assigned, erroring, restarting and repeating.

Start with the Pulsar source's checkpoint handling and pipeline restart path described in the issue. Reproduce or inspect behavior with partitioned topics and the 1s failover timeout; done means partition reassignment no longer causes an error/restart loop while preserving exactly-once semantics.