Gpu min max, more pattern implemented

Presentation

This mean the not available case 1010, 1011 and 0111. The following case where implemented with reshape for min/max, but not for sum:110, 001, 0011, 100.

Also add test that other GPU min/max that are done by reshape.

This probably fix a the bug in the not enabled implementation: 010_inner

Fix our algo to find MKL with Canopy on windows and linux

[WIP] fix gh-1344. Scan opt error/warning.

Add theano.sparse_grad interface.

Bouchnic truedot rebase

Add documentaiton about the MPI and load ops

Fix test that fail depending of the seed