Fix memory leak in GpuCAReduceCPY

don't move op that don't support float16 on the GPU

(crash fix) Fix availability of cuDNN enumerations in v5 vs v6.

Fix arguments order in register_specify_shape_c_code of tensor/type.py

Wrap Op params for many gpuarray DNN Ops and add cuDNN v6 integration.

dockerfile for jenkins buildbot

add magma to buildbots

added description to strides in pools

Correctly handle offset in GpuCumOp

added an example to make it easy

Ops rewritten:

 - GpuDnnConvDesc
 - GpuDnnPool
 - GpuDnnPoolGrad
 - GpuDnnConv
 - GpuDnnConvGradW
 - GpuDnnConvGradI
 - GpuDnnBatchNormInference
 - GpuDnnBatchNormGrad

cuDNN v6 integration:

 - Support MAX DETERMINISTIC algorithm for GpuDnnPool with cuDNN v6.
 - Update pooling tests for DNN module so that they use the right available algorithms depending on runtime cuDNN version.
 - Allow CPU Pool and PoolGrad ops to use MAX_DETERMINISTIC algo when cuDNN v6 is used with GPU counterparts.
 - Encapsulate cuDNN constants used in DNN module, to help choose the right cuDNN definitions depending on the runtime cuDNN version.
   Currently supported cuDNN versions: v5.1, v6.0.

Do not forget offset in GpuCumOp

crash fix and opt warning fix(make opt apply)

outdim -> ndim leftover and the same to is_flat

Wrap Op params for theano.tensor.nnet.corr.BaseCorrMM and sub-classes.

Fix the opt crash regression introduced by gh-5950.