Summary:
On Windows we have to retry the delete a couple of times.
The reason for this is that a `FileDelete` command just marks a
file for deletion. The file is really only removed when the last
handle to the file is closed. Unfortunately there are a lot of
system services that can have a file temporarily opened using a shared
readonly lock, such as the built in AV and search indexer.

We can't really guarantee that these are all off, so what we can do is
whenever after a `rmtree` the folder still exists to try again and wait a bit.

Based on what I've seen from the tests on CI server, is that this is relatively rare.
So overall we won't be retrying a lot. If after a reasonable amount of time the folder is
still locked then abort the current test by throwing an exception, this so it won't fail
with an even more cryptic error.

The issue is that these services often open a file using `FILE_SHARE_DELETE` permissions.
So they can seemingly be removed, and for most intended purposes they are, but recreating
the file with the same name will fail as the FS will prevent data loss.

The MSDN docs for `DeleteFile` says:

```
The DeleteFile function marks a file for deletion on close.
Therefore, the file deletion does not occur until the last handle
to the file is closed. Subsequent calls to CreateFile to open the
file fail with ERROR_ACCESS_DENIED.
```

Retrying seems to be a common pattern, SQLite has it in their driver
http://www.sqlite.org/src/info/89f1848d7f

The only way to avoid this is to run each way of a test in it's own folder.
This would also have the added bonus of increased parallelism.

Reviewers: austin, bgamari

Reviewed By: bgamari

Subscribers: thomie, #ghc_windows_task_force

Differential Revision: https://phabricator.haskell.org/D2936

GHC Trac Issues: #12661, #13162