Problems to fix:
- dnn pool grad with ignore_border=False do not seam to compute the good value. Did we have a bug? Do they support it? Maybe not as it isn't documented.
- We do not tests mode=average correctly as we do not have a reference implementation.
- We do not tests pad. Need a reference implementation.

Fixes and cleanup for maxpool with padding

Check negative strides in gpu_contiguous

Escape backslashes in windows path string

Fix opt crash due to new CPU parameter

support max pooling with padding

[WIP] Detect Null gradient wrt non-sequences in scan

Additional pep8 fixes

Alter test case to avoid use of advanced indexing

[CRASH] Error when scan has only nonseq inputs

Crash fix

Speed up local useless subtensor

Enable algorithm selection in cudnn convolution

Fix cudnn crash with v1