Python2's weird comparison semantics are inconsistent with what
we expect when `NotImplemented` is returned.

If the arrays do not have the same shape then we return NotImplemented.

A new method `object_size` is added to the Array template class, which
calculates the total size of the Python object in bytes.  The function
`size_of` then wraps this method and provides the necessary interface
for a Python method object.

This makes it fail for Python 3 also on 64 bit platforms.

The motivation behind this change is to allow keeping Tinyarray
distributions in git repositories that have a different layout from the
canonical Tinyarray git repository.  This is useful for example for
packaging.

The static version is now kept in the file "version" and can take a
special value that signifies that the version should be determined using
git.  This file is rewritten upon creation of a source distribution.

Also, be explicit about Python 2 support. Drop development
status to 4 - Beta until we can test on other platforms.

In Python 3 changes the semantics of the `/` operator for integers.
See https://docs.python.org/3/howto/pyporting.html#division
for details. This required that we convert Array<long> to Array<float>
before doing the division

There were several changes in hashing between Python 2 and 3,
including several macros being added to avoid "magic numbers".
In addition, instead of having a hand-rolled implementation of
hashing for `double` values we just use the Python API function
`_Py_HashDouble` for compatibility.

There is no `PyUnicode_AsString` in Python 3, so I have no idea
why this even compiled.

Ran 2to3, removed conditional cPickle import and replaced `div`
operation with `truediv`.