Was:

__remove_clients__ -> __prune_r__

inline __prune_r__ inside __remove_clients__

This should help gh-3341 and speed up a little compilation of big graph.

The recursion was:
__remove_clients__ -> __prune_r__ -> __prune__

This help gh-3341

This should also speed up change to the graph, but I didn't do any timming.

[ENH] Keep an history of only 1 checkpoint of FunctionGraph

[ENH] Speed up bn high_mem, infer_shape, trac more stack trace.

Small tutorial for op params

Fix unhashable type if shapes are list

Fix return value of R_op in HostFromGpu/GpuFromHost

Update .travis.yml

Update install path of miniconda

Fix BaseAbstractConv2d to convert kernel_shape and image_shape to tuples

Fix a crash when using Softmax with the CLinker

Implemented grad for Images2Neibs using scan.

Correct spelling mistake.

Fix unpickling of Gemm and GpuGemm Ops

Fix infer_shape of Join when axis < 0

Fix extra scan due to old information in infer shape.

Fix index errors

Python 3 doesn't have numpy packages with openblas, apparently.

Re-enable mkl for python 2.6 since then aren't numpy pacakges of 1.7 with openblas.

[CRASH] when Elemwise have more then 31 inputs

fix for aarch64 (ARM 64-bit arch)