The unicode tests would often fail when tested in parallel because
we were not exporting to unique file names. From what I understand,
a variant similar to the following race condition occurred:
1. Thread A exports to file blah.pdf.
2. Thread B exports to file blah.pdf.
3. Thread A confirms file blah.pdf exists.
4. Thread A deletes exported file blah.pdf to clean up.
5. Thread B fails to find file blah.pdf and reports a failure.