Disable the optimization to use DnnSoftmaxGrad for cuDNN v3 since it …

Disable the optimization to use DnnSoftmaxGrad for cuDNN v3 since it is broken for (n, c 1, 1) shapes at the moment.

Average pooling

Fix buildbot

Make broadcast fixup in filter_variable() depend on an argument and stop doing it for shared variables.

The inner loop that would update the broadcastable pattern according
to the rebroadcast dict would overwrite the idx extracted from the
Subtensor, thus always selecting the last dim.