Parameterized local_elemwise_alloc_opt for GPU support

Cudnn2

Conflicts:
	theano/sandbox/cuda/tests/test_dnn.py

Conflicts:
	theano/sandbox/cuda/dnn.py

Fixed #2071: Moving test_record.py from Pylearn2 to Theano.

timing information added in debugprint

Fix scan handling of error when on CVM.

[BUG]fix gh-2348

Cudnn r2

DownsampleFactorMax support strides: issue #2196

[CRASH] Fix scan unpickle (crash) as we added a new attribute. Also make a scan ...

Fix scan unpickle (crash) as we added a new attribute. Also make a scan opt support old pickled Subtensor.

 - Split the C code apart in between the Conv ops (the previous
   solution is not workable with the new interface).

 - Move the C code out into separate files using COp.

 - Make cudnn_helper.h provide wrappers exposing the R2 api with the
   R1 functions when possible to reduce #ifdef maze in actual function
   code.

when constant are used, we use the smallest dtype for them. But this
isn't enought when computing on them. So upcast to int64 to make sure
compuration is right.