Create GpuErfinv and GpuErfcinv. Make sure
these GPU ops return the same values as
CPU (scipy.special) equivalent ops.

Update optimization to take account of these ops.

Optimization is applied only with CUDA device

(a warning message is printed if OpenCL is used).

Add tests.

Remove the custom version check in the __init__ of pygpu.

Moving ExtractDiag op to the new backend

fix crash in MRG_RandomStreams with the new backend

Fix reload pickle

DestroyHandler should not overwrite existing rval[app] values

[CRASH] fix crash. fix gh-5291.

This solves a problem for Ops with more than one in-place input.

broadcastable test for binomial, multinomial and normal

Uninitialized memory returned from empty GpuGemv dot-product

Fix in advanced_inc_subtensor1 if index is broadcastable

Gpuarray average and max pooling

Clean up and also fix the crash in the code example of gh-4679.

Removed save_every_N sequence size limitiation.

Fix blas detection in corner case

Clarify GPU memory pre-allocation in new and old backend