Previously, several tests in FileSystemWatcherTests assumed that all
of the changes detected by the watcher would be grouped into a single
change set. This assumption breaks down when a test runs slowly (due
to CPU or IO contention, for example), and making changes to the file
system takes long then the watcher's polling interval. When this
happens, the changes will be split across two (or more).

This commit attempts to make the tests more robust. The tests now
tolerate multiple changes sets by combining them and asserting that
across the n change sets, only the expected changes were detected.

Closes gh-25901