merge: flush any deferred writes before, and after, running any workers
Since we fork to create workers, any changes they queue up will be lost after
the worker terminates, so the easiest solution is to have each worker flush
the writes they accumulate--we are close to the end of the merge in any case.
To prevent duplicated writes, we also have the master processs flush before
forking.
In an in-memory merge (M2), we'll instead disable the use of workers.
Differential Revision: https://phab.mercurial-scm.org/D628