Move the autoselection variables to be per-thunk rather than per-program

Compatibility between both GPU back-end and doc.

Relu

Move the autoselection variables to be per-thunk rather than per-program since that's what it should have been in the first place.

Small stuff

Doc Ubuntu 16.04 and clean up flag in_c_key

[CRASH, ENH] Ifelse grad and speed up perform

Fix crash and add cuDNN in version to help not delete cache when updating cuDNN.

crash fix gh-4460

Small doc fixes

Fix slips from PR #4244

Tested on Ubuntu 16.04. On first two runs the result is 0.71 or 0.7,
but then it is consistently 0.66s.

Vectorize MultinomialFromUniform.perform() to speed up tests in FAST_COMPILE/DEBUG_MODE. This cause some failures due to no output for 10m on travis

flake8 sandbox/cuda/*.py