Fix scan pushout

When running on GPU, and constructing the list of inputs, they might not end
up on the right device (i.e. the input is on gpu instead of cpu or the other
way around). The simple fix is just to filter the arguments.

Get the lock before compiling cuda_ndarray.so file.

fix __getitem__ on shared sparse variable and test it.

No change in NEWS.txt as this was not released.

Bug fix in Scan

Fix test

use smaller shape to prevent DebugMode to return error sometimes due to different rounding on CPU and GPU with the Sum op.

New elemwise

Added shape_of_variables function

Given an env and the shape of all inputs returns a dictionary holding the
shapes of all intermediate and output variables

Load op

Make ifelse work on gpu

Also moved relevant test

Enable fusion of node that have the same clients multiple time.

Fix new op

Catch more gpu init error.

Two extra op from numpy.