Increase NUM_THREADS

[bendavid]

Modern dual-socket servers are commonly available with up to 384 threads now (and soon 512 threads with EPYC Bergamo).

Suggest increasing NUM_THREADS from 64 to even 512 (especially for the sake of default numpy and scipy wheels)

[mattip]

This would negatively impact limited systems, since a buffer is pre-allocoated for each thread. People running on such server systems are more likely to build their own software from source and not use the PyPI wheels.