Negative axis were automatically handled by numpy implementation inside
Join.{perform,c_code}. However, both Join.grad and
GpuJoin.{perform,c_code,grad} implementations were not thought to take
care of negative axis.

This commit adds checks for exceeding negative axis in Join.make_node for
constants and in {Gpu}Join.{perform, c_code} for symbolic variables. Axis are
converted to positive values inside GpuJoin.{perform, c_code} since current
implementation does not support negative axis.

Fix use of tabs in python file

[CRASH] Improve shared variable management in Scan

Implemented fix for issue #2613: Concatenation bug for negative axes on ...

WIP: Custom pickler for shared variables

[BUG, TEST] Fix bug in numpy 1.9 and test in DebugMode

Binary mode for test files

Added import

Reading in binary mode

Used BytesIO

Binary mode for zip

Binary mode in tests

FLAKE8

Added test pkl_utils into flake8 whitelist

Refactored pkl tools and tests

Removed test_pkl_utils from flake whitelist

Use doctest to update pydotprint plots and debugprint output

Average pool CPU with python code

[CRASH] Make MaxAndArgmax support NoneConst as axis.

[TESTS] Fix crash in tests in debugmode and small optimizations.

Merge conv2

Added doc for strict flag.

Small doc fix gh-2793

Change cache refresh patter for speed up