First attempt at fixing issue 2454: making consider_constant() functiona...

fix gh-2581

Convert inputs of maxpool to tensor

Correctly handle gradient of identity on int

Flatten C code

[BUG] pool with padding

do not try cudnn opt when it isn't available

Included changes in tensor/opt.py for #2454. These were previously lost due to problems committing to GitHub.

Set borrow=False for scan inner inputs

Add missing `host_from_gpu` import to `cuda.opt_util`

[WIP] Mechanism for running a test multiple times with different random seeds

Allow dnn convolution to work inplace

Update aa.py, theano.sandbox.wraplinker seems not exist.

Fix few tests for the nightly buildbot

Dnn pool support of pad