Speed up some tests

Try to fix travis. Don't recompile du to new flag when not needed.

Bump min Python version to 3.4 (fix slowness travis in Python 3.3)

Add a DEBUG option for C code

Add deterministic={default,more} flag

Params for other ops

Enable float16 for GpuGemmBatch.

Fix opt crash in float16 and enable C code for MaxAndArgmax and Argmax

Improve usage of op params for neighbours ops (try to fix jenkins error).
Update AdvancedIncSubtensor1.c_code_cache_version (as we change the type of an op param).

Add reduction based on cudnn.

 - 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)