Skip test_sync_update in DEBUG_MODE because it is incompatible

Broken link in doc to amdlibm

Fix scan MissingInputError cause by hidden var used in condition.

Doc magma

Don't ask for MKL_THREADING_LAYER=GNU if NumPy is fixed

Follow API changes in cudnn 7.1 for RNN params

Also default to clang for FreeBSD.

Try cygwin fix

Fix scan names of variables.

Support more recent flake8 version

After looking at older documentation of flake8, I think all flake8 3.* version should work: http://flake8.pycqa.org/en/3.0.0/user/python-api.html. So removing the version check completly.

Fix gh-6566. Update the pictures in the convolution tutorial.

Better error msg

Fix non-cudnn reductions that were all subtly broken by 91bc16c3.

support allow_remove_{inf,nan}