Executing the libc GPU tests multithreaded may overload the GPU runtimes
for CUDA or HSA, which results in test errors. In addition, libc seems
to not respect the -j option of ninja, when it is invoked at the top
level.
This patch adds special handling of the check-libc rule when passed as
additional check to the OpenMP builder. Ninja will thus first cd into
the runtimes build directory to make the libc testing respect the given
amount of parallelism.
We then limit the number of threads to 1 as to minimize the potential
for any errors with GPU runtimes.