sklearn crashes when using cross-validation algorithms due to small /dev/shm

[agude]

sklearn will blow up when running CV (cross-validation) algorithms if you try to use multiple jobs because /dev/shm is too small. See:

• joblib breaks automatic memory mapping with large arrays: multiprocessing pool self.value out of range for 'i' format code scikit-learn/scikit-learn#3313
• Detect when /dev/shm is size limited joblib/joblib#168

We can either increase the size of /dev/shm (although it would have to go from O(mb) to ... much bigger) or we could set a new location fro sklearn with JOBLIB_TEMP_FOLDER. I think the solution should be implemented on our side instead of the user side for consistency.