Summary:
Package registration does not seem to be thread-safe on
Windows.  Placing the system under heavily load seems to
trigger registration failures even though they are all
different package-dbs.   This makes the plugin tests
a bit flaky.

I think this is because on Windows we use pessimistic locks
while on Linux we use atomic file replacement.

On Windows ReplaceFile is atomic, just the metadata write
may not be.  Since the metadata is not of importance
we should either switch over to ReplaceFile or
fix the locking code to not error out but wait.

For now however I have to force these 25 tests to run
serially in order to guarantee their correctness.

Test Plan: ./validate

Reviewers: bgamari

Subscribers: rwbarton, thomie, carter

GHC Trac Issues: #15313, #13194

Differential Revision: https://phabricator.haskell.org/D4918