Prepare release 0.9.0rc3.

Last before rc3

The predefined constants (enum) are the same with v6

Fixes #5655

Change the default of conv.assert_shape flags to speed up compilation time

Faster compilation and import with cuda backend

Remove old files not used anymore.

A couple of test fixes

Fix problems in daily buildbot (debug mode)

Warn if taking softmax over broadcastable dimension

We already encountered then in ProfileStats.__init__

The gpu back-end may have been imported, but we are not even timing a GPU function

Otherwise it is not actually testing anything.

Fix precision in cudnn

always the product of a matrix with its transpose.

Code simplification. The function are the same except for the default value of the subsample parameter.