Import numpy

[REG, CRASH] fix crash related to float16 introduced in gh-5077

warn when profiling.ignore_first_call isn't used.

Make test for cast to/from int/float16 and make sure to have a cast in the inner graph instead of identify. (it was still working with identity, but having an explicit cast is better)

Fix cache

Fix crash introduced by gh-5077 by keeping GpuElemwise as original and only casting to float16 in the code generator.

Add a method to properly convert the graph of Composite to float32.

Gpuarray pool grad grad

Fix some failing tests in debugmode

[REG,BUG] Use _props_dict() from the right Op.

CudaNdarray.filter rejects broadcastable dimensions with non-zero
strides

Add Max3dPoolGradGrad which accept 5d input and pools over last 3
dimensions. Also add padding parameter and now accept tensor variables
to make_node which removes necessity to check for constant args. Local
optimization are updated accordingly which now work for 2d and 3d
pooling.

It was using the one from GpuFromHost instead of AdvancedIncSubtensor1.