Refactor commit logic (#2580)

* Add tests for commit consistency showing wrong behaviour

* Clear scheduled document changes at the end of a commit operation

* Rename private variables to communicate intent

* Remove obsolete comment

* Use different error code

* Use single mongos in consistency tests

This ensures that the fail points are created on the same server that the write operations take place on, which can't be guaranteed in a sharded cluster with multiple mongoses.

* Rename test methods for clarity

* Explain reasoning for index error

* Extract helper method to create failpoint