Be compatible with Numpy 1.16

MAINT: collections import is deprecated

GpuContiguous need a perform method.

Fix test in FAST_COMPILE mode. GpuContiguous need a perform method. This should fix test in the daily buildbot

Fix syntax error in format string in gpuarray/multinomial.py

Update extension suffix config var name

This should remove the deprecation warning on Python 3 where the configuration variable has been renamed to `EXT_SUFFIX` from `SO`.

float64 support for GpuCholesky, GpuCusolverSolve, GpuCublasTriangularSolve and their L_op

Better error message to help in case like gh-6639

* Add gamma family functions from SciPy with C code for GPU

The C code is MIT-licensed.

* Remove commented code and bring in missing functions

* Indicate scalar definitions for Elementwise extension

* Upgrade to float64 as previously

* Add docstrings and remove chi2sf duplicate

* Added support for the axis parameter of numpy.unique, which became available since numpy.__version__ 1.3.0.

* Fixed doc example. Also fixed typo of the numpy version which added the axis parameter. The version that added axis was 1.13.0, not 1.3.0.

* Made setup.py, conda and travis require numpy >= 1.13.0 in order to get numpy.unique with axis support. Changed test_Unique to test for axis=0 and axis=-1 parameter values. Fixed Unique.infer_shape bugs when axis is not None.

* Updating scipy version to 0.19.1

* Updated miniconda version.

* Fixed Unique Op doc typo. Split tests for Unique axis parameter. Success or failure now depends on numpy version. Added numpy version 1.13.0 and 1.9.1 as test cases in travisCI. Reverted required numpy version down to 1.9.1 in setup.py and conda.

* Fixed Miniconda version to Miniconda2. Changed python3.6 to 3.7 in travisCI tests.

* Fixed Unique Op doc for non axis handling. Changed python 3.7 to 3.6 in travisCI.

* Added export MKL_THREADING_LAYER=GNU to travis script. This seems to be necessary because the updated miniconda2-4.5.11 installs 2018 MKL.

* Fixed doc example when numpy version is < 1.13.0. Changed python 2.7 environment's scipy and libgfortran versions to avoid a build conflict in travisCI. Changed conda python 3.6 environment versions to avoid a build conflict in travisCI.

* Updated numpy 1.13.0 to 1.13.1 in order to have numpy built with mkl support.

* Fixed flake8 formating error.

* Removed Unique examples in doc with axis input. These raised errors with sphinx, and conditional skiping was added in sphinx1.8, which is a higher version than the ones we test with.

* Changed test_basic::test_grad `class O` to `class Obj1` to try to prevent flake8 ambiguous class definition E742 format error in python3.6 environment.

* Hoping this makes sphinx not interpret 'py_' as a target name, which it does not know...

* Changed Sphinx version down to 1.5.1. I had mistakenly thought that sphinx was responsible for downgrading numpy not to use mkl.

* Fixed bug in tests. Instead of passing a function object to assertRaises, I was evaluating the function and passing its result.

* Reduced the ammount of travis builds.

* Removed trailing `<<: *normaltest` from .travis.yml.

* All tests in `test_Unique_axis`, except `test_op`, now raise a `SkipTest` exception when the numpy version is lower than 1.13

* Implemented @nouiz's suggestions.

Modifications for 1.0.3 release

Broadcast sparse dot

Python 3.7 compatibility: Regenerate scan_perform.c with more recent cython.