- use KERNEL macro
- do not use `%(fail)s` on GPU to avoid returning prematurely from
  kernel
- have special block for y == 0 (and reorder other ones)
- keep calling // 0 or % 0 on GPU, even though cuda will not fail

Even if the output shape is incorrect, we should not crash.

Raise a Python exception for integer division by 0

Unsigned char:  fix: theano/scalar/basic.py

change default dnn path on windows

Changes to GpuEye to enable `k` offset parameter.

Remove call to GpuKernel_init with binary

As it is trying over and over to compile it as source.
Also, start removing generation of binaries from Theano.

Disk cache option for gpuarray

Implement new op param type `EnumType`

preallocate with a direct call.

Updated numpy as np #4218

Fix memory leak

Remove the old backend and all of its tentacles