Multi-gpu support

test function is moved to another function

Grad argsort

Fix dnn_conv and dnn_pool docs

"each of" the top and bottom, left and right, etc.

tensor ones and make GpuGer conform.

Also fix the logic in dnn_available() so that we always have an error
message to present.

Finally add the context name that failed in the NoCuDNNRaise error message.

Convert double recursion to simple recursion

Link d3viz doc in main tutorial

Fix GpuFromHost compilation error using libgpuarray

Scan fixes and speedup

Fix daily buildbot tests