- i (scalar, int64)

 - mode (enum list)
 - context (gpu_context_type)

Add Kernel param for theano.gpuarray.neighbours.GpuImages2Neibs:
 - mode (ga_int)

That helps very much to reduce C code variability
(from 18 individual ops to about 6 modules with more than 1 op).

 - mode (enum list)

 - inplace (bool scalar)
 - set_instead_of_inc (bool scalar)
 - context (gpu_context_type)
 - ndim_input_0 (size_t scalar)
 - ndim_input_1 (size_t scalar)
 - typecode_input_0 (integer scalar)
 - typecode_input_1 (integer scalar)

The 4 last params are used to reduce C code variability into c_init_code_struct().

Wrap Op params for theano.gpuarray.subtensor.GpuAdvancedIncSubtensor1_dev20:
Same as theano.gpuarray.subtensor.GpuAdvancedIncSubtensor1.

 - inplace (bool)
 - set_instead_of_inc (bool, converted to int)

Set `check_input=False` for theano.tensor.subtensor.AdvancedSubtensor1.

 - inplace: boolean scalar
 - ndim: integer scalr
 - otypenum: integer scalar
 - otype_is_float32: boolean scalar
 - otypecode: integer scalar
 - context: gpu_context_type

 - context (gpu_context_type)
 - axis (integer scalar)
 - device_is_not_cuda (bool scalar)

speed up tests, fix compiledir cache crash and old unpickling.

Added the tri-gamma function.

Params for Gemm, CGer, CGemv

Gradient for matrix pseudoinverse

Work around erroneous dnn_conv results with algo small an large batches

Buildbot testsuite names

removed part of _optcheck_fgraph and called std_fgraph instead

(replace algo `small` by `none` if necessary).

No need of big input.

Allow fallback on GpuCorrMM if cuDNN is not there