changed CLinker.cmodule_key to do the real computation in a static method to allow calling it from elsewhere.

CLinker returns keys that include the role of each variable, as well as the
no_recycling parameter.

This fixes an incorrect result I am seeing, and I believe it fixes the
mysterious 'output-borrowing' behaviour that Olivier Delalleau reported.

FAST_RUN to the behaviour of profile mode.

added a test of softmax optimizations to make sure that large positive and negative arguments overflow gracefully

code.