Improve support of MSK rolling upgrade

Since NXP-30709 there is a retry mechanism if the computation thread is not able to checkpoint its position (Kafka is not available because of a rolling upgrade).
Nevertheless after 3 checkpoint failures the computation thread gives up and terminates.

The problem with the NXP-30709 implementation is that the checkpoint failure counter is not reset, and there is no more retry on the third rolling upgrade upgrade.

We want to terminate only after 3 failures on the same checkpoint.