Disabling OpenMP parallel pragma for CPU tensors causes performance regression

[alextnewman]

The removal of OpenMP from this tensor_cpu_inl.h caused a massive performance regression for us on Windows (MSVC 2013), Mac (Clang), and Linux (gcc): f225763

Locally, we've reverted this commit and gotten a tremendously positive result (20%+ improvement in training time), so it would be very helpful if there were some sort of option or flag we could use to enable OpenMP parallelization for this function without internal forking.

[szha]

This code base has been donated to the Apache MXNet project per #373, and repo is deprecated. Future development and issue tracking should continue in Apache MXNet.