Perf drop with OMP_NUM_THREAD/MKL_NUM_THREADS

Thanks to @mavlyutovr's investigation we found out that the throughput of our TBB-based data pipeline API can drop significantly if OMP_NUM_THREADS and/or MKL_NUM_THREADS are not bound to a low thread count. We should investigate if there is anyway for us to mitigate this issue on our end. If not, we should document this behavior and advise users to be cautious about it.