The commit cleans up the code, such that it compiles without any warnings.
Additionally it is ensured that no segmentation faults occurs. There have
been several different types of changes.

 - Unused variables have been removed.
 - A lot of cleanup and documentation has been added in the course of issue #2.
 - A numpy deprecation flag for version 1.12 has been set.
 - Attention has been paid to the difference between PyArrayObject and PyObject
   pointers.
 - Checks to detect memory allocation errors have been added, which would have
   otherwise resulted in segmentation faults.

Closes #7.

Resolve "Fix reference counting"

Closes #2

See merge request !4

Remove the print statements used to debug reference counting issues.

Perform a code review to check that the reference counter of the input arrays
are not modified by the function alls and that there are no lost references to
the output array.

This closes #2.

Resolve "Create benchmarks"

Closes #1

See merge request !3

The default runner has not enough memory to execute the benchmark job. Therefore
a specialized high-memory runner should be used.

Change path to paths

Change artifact to artifacts

Implement a benchmark script which compares the execution times to numpy and
generates comparison plots. The benchmark is executed in a CI job. The plots
from the CI job are embedded in the README.

Resolve "Create unittests"

Closes #4

See merge request !2

Add the missing __init__.py file which promotes the tests directory to a python
package. This solved the unit test execution issue.

This reverts commit 5d430d59. Package is already
installed due to 'before_script'.

Test whether installing (and compiling) the package before executing the unit
tests resolved the loader issues.

Change state to stage.

Raise an exception if any of the arguments id multi-dimensional. See also #2.

Use PyArg_ParseType's O! format and require PyArray_Type for both arguments, to
ensure that the passed object is an array. If this test fails, it raises an
TypeEror exception.

The added unit testes fail currently. There is an issue with the exception
handling if invalid arguments are passed to the methods (see #2).

This closes #4.

Ci

See merge request !1

Building the extension requires numpy to be installed, therefore it is not
sufficient to specify as a dependency. Numpy has to be install before hand.

Fix accidentally wrong file extension.

Add a basic README with installation and usage information. Use doxec to test
the information in the document.

First implementation to intersect two sorted numpy arrays. This completes the
feature list of the experimental implementations. Further improvements can be
achieved by using a binary search or a galloping search.

First implementation to merge two sorted numpy arrays. The purpose of this
version is to learn Python's C-API and Numpy's array API. This version is not
stable and works only for 64 bit floats. Merging a 35M and a 15M array takes
about 10s on my laptop.