[DON'T MERGE] Supports Argobots

[shintaro-iwasaki]

It works in the following environment (MPICH's Argobots example (ult_send_recv.c))

• OpenMPI libraries with --with-threads=argobots with and without --enable-debug.
• OpenMPI-master Pthreads-version mpiexec
• OpenMPI-master Pthreads-version mpicc
• Argobots

It can run OSU benchmarks as well, but it is extremely slow (10~100x, need to be extremely patient)
To test on my laptop (Ubuntu 16.04 LTS), I need to pass --mca-btl self,tcp to mpiexec; otherwise, it hangs.

Pthreads version also works (make check and a few OSU benchmarks as far as I tested).

• OpenMPI libraries and binraries with --with-threads=pthreads without --enable-debug.

All the update should be included in #4, but to run make check, you might need to modify test/*/Makefile.am.

We need some features in the future:

- opal_yield()
- opal_usleep()
- mca_close_component() for a threading library (there's no way to finish a threading library)
- preprocessor for ULT to control msecs of (e)poll()

[shintaro-iwasaki]

I found the cause of significant performance degradation; I by mistake set the polling interval to 1 [msec], not 0. The latest version shows much better performance. On my laptop, the debug build shows the following performance:

# OSU MPI Bandwidth Test v5.5
# Size      Bandwidth (MB/s)
#                 Before     After
1                   0.01       1.81
4096               27.25    1568.40
4194304          6412.28    7199.88