Incorrect bookkeeping of outstanding produce requests on produce failures

[gabrielreid]

Versions used

alpakka-kafka 2.0.1

Expected Behavior

When an alternative supervision strategy is used with Producer.committableSink, a producer failure should just be ignored and not have further impact on the behavior of the stream (i.e. the stream should still commit outstanding offsets when stopped).

Actual Behavior

The internal bookkeeping for the number of outstanding produce requests is not correctly updated on a Producer failure. This results in no automatic commit occurring when the source of the stream completes, meaning the stream will stay open until a time-based commit occurs (which could be a very long time, depending on the commit settings)

I noticed this while going through related code while investigating the cause of #1040

I'll add a PR with a test that demonstrates the failure and a fix.

[ennru]

Fixed with #1043