Crash in scikit-learn RandomSearch with n_jobs=8

[alexjc]

Theano's compilation compile cache is locked for a single thread, but execution should work with multiprocessing to help max out the CPU usage. However, currently Python process crash randomly when using n_jobs=8.

[alexjc]

It's possible to set different cache directories for Theano, which would duplicate the compilation directories, but potentially speed up the process if recompilation is done often.