threadsafe job counts; configurable batch_size

[gojomo]

Fix for #571. Uses local vars in each thread for counts. Adds class/init batch_words parameter to set a non-default value.

[gojomo]

@tmylk @piskvorky – any concerns about merging this into develop? I believe it will improve (and perhaps fix entirely) two causes of intermittent CI test failures (and thus save anyone else working on develop from thinking their changes may have caused a failure, as in #538).

[piskvorky]

It looks quite elegant, I'm not sure why I messed around with shared state when a simple counter was enough.

Maybe I'm missing something here (reasoning about parallelism is tricky) but if all tests and sanity checks pass, it's LGTM from me (+ the obligatory CHANGELOG entry).

[piskvorky]

@tmylk I suspect a similar issue could be behind the failing tests in LdaMulticore (which uses processes instead of threads, but otherwise similar logic flow, AFAIR).

[gojomo]

Added CHANGELOG & merged to make it easier for other in-progress/future branches to build on this.