GitLab

- doesn't work right yet

- note: there isn't really any harm in creating many ULTs, but we still
  throttle for consistency with the pthread version of this test and
  also to make sure the I/O phase doesn't result in too many open file
  descriptors

- io buffers can be allocated per thread as before, or with a single
  shared buffer for all threads

- thread_join appears to cause a busy spin, which makes the main thread
  interfere with the remainder of the benchmark even when using
  abt-snoozer