reduce number of WrongValue errors.

to help reduce number of WrongValue errors.

when they are chosen at runtime.

Fix typos.

given data type configuration for given ndim.

Fix test cases generators to avoid cases that are intended to fail.
Fix messages and properly exit from CUDA in cuDNN conv ops.

Update cases for gradweight ff_tiling.
Allow DOUBLE_CONFIG for all fft_tiling computations in cuDNN V6.

Add a script to quickly run specific tests.

Add a first special test case for conv 2d fwd.

Update interface of test-cases generator.
Reduce number of generated test cases.

Cancel changes into dnn_fwd.c
Heavy simplification of check_dnn.
Make check_dnn runnable as a python script.

- add sub-module check_dnn_doc.py to help check if tests
  fail as expected from cuDNN documentation.
- re-organize whole code into `check_dnn.py`.
- add a test case generator.
- add classes to help choose algos at runtime without actually
  computing convolution.
- Separate exhaustive tests and tests for choosing algos at runtime.

with list of deterministic algorithms
and other utility definitions.

that take algo, alpha and beta as parameters.

Rewrite tests to handle both 2D and 3D conv correctly.
Simplify gradinput and gradweight tests.

Synchronize configdefaults with #5964.

Disallow float16 precision for convolution gradients functions.

Use the right mask to check for alignement.

Make setup.py fallback when versioneer can't get the version.

ARange gradient (contd)