Parallelization of the test suite

[Clement Pernet]

Hi,

Many files in the test suite now use parallel task here and there to speed up the computation:
mostly the verification phases, which use a lot of fgemm like e.g. in test-lu, test-echelon, etc.

As a consequence, running the test suite in parallel, with make -j, is extremely slow: presumably, p
processess each running p threads is not a good option on p processors!

Two options to avoid it:
1/ only run the test suite without the -j option
2/ remove the thread parallelism in each test file

Option 1 has the drawback that the compilation is now only run by 1 proc, which slows things down
by quite a bit.

Also, I compared the time of running the executables in the two options (not counting the
compilation time) on my 2 core laptop
Option 1: 1m2.210s real time with NUMTHREADS=2 1m9.239s with NUMTHREADS=4
Option 2: 0m34.807s real time with make -j 2

Hence I suggest to remove the parallelization within the test files for the moment. If they need to
be parallelized, it will more likely be on the outer loops (all test instances) rather than in the
linear algebra dimensions of the problem, which are too small to hope for a reasonable speed-up.

Clément