Handling of `nthreads` about fit-sne

[linqiaozhi]

1. It was implemented this way while ago in #6, and I never got around to changing it. You're right though, there is no reason why this parameter should passed differently than the others.

2. and 3. Fixed in #41. This was a particularly important catch, as the computeError() were not multithreading at all, as you reported in #37. On my machine with 12 cores (2 threads each), this takes the compute error from about 15 seconds to 1 second, as it parallelizes very nicely.

3. After #41, the only loops that do not use PARALLEL_FOR are the ones where we have a progress bar that should only output on a single thread (it checks if t==0). I don't know how to do that in a way that works for both OpenMP and C++11 threads, so I just hardcode it as C++11 threads for now. Should be changed in the future though.

Thanks @dkobak!

[dkobak]

OK great! I think (1) is not important and it can stay like that. I close this.