Ensure that blas_thread_init has been called in openblas_set_num_threads

[embray]

Not doing so can lead to bugs if the process happens to have been forked at some point. For example:

pid = fork();
if (pid) {
    openblas_set_num_thread(8);
    // <call some multi-threaded blas function (dgemm, etc.)>
}

This leads to segfaults because currently openblas_set_num_threads assumes blas_thread_init() has already been called and blas_server_avail == 1. But after a fork, due to the pthread_atfork handler which calls blas_thread_shutdown, this is no longer the case. But then when you get into exec_blas_async it does call blas_thread_init() even though there are already threads running, creating a likelihood that one or more of those threads segfault.

[brada4]

Should work fine.
It is one-time code in general, "unlikely" tagging just impairs readability.

[embray]

Sure. It was literally just copy/pasted from elsewhere. I see no harm in including unlikely() but I also agree that in most cases it is not significant.