parallel chunking routine and Python 2GB bottleneck for large statsmodels OLS

[turbach]

Appears to be a monkey jar ... the chunker ships jobs out to the pool that fit through the 2GB bottleneck then statsmodel OLS fitting inflates the size and the returns are too big to make it back through the bottleneck.

Possible workaround: use tester fit to estimate OLS return size and chunk accordingly for 2GB, fall back to serial.