Fixed memory leak in contour as reported in#6940. PR is based on v2.x branch, but fix needs to be in all active branches.

Problem is down to my stupidity when adding numpy arrays to python lists to return from contour/contourf calls. New contouring C++ code uses our numpyarray_view wrappers, which are in charge of their own reference counting, and native Python/C API list functions for which I am explicitly responsible for reference counting. Consider code:

numpy::array_view<double, 2> line(dims);   // ref count 1PyObject* pyline = line.pyobj();  // ref count 2PyList_SetItem(pylist, pyline);  // ref count remains 2 as PyList_SetItem steals referencereturn pylist;  // line goes out of scope so ref count 1 for the one and only reference in the list

This is fine, but I am usingPyList_Append instead ofPyList_SetItem as I don't know the length of the list when it is created. It turns out thatPyList_Append increments the reference count whereasPyList_SetItem steals a reference. I assume I used to know this but have since forgotten it and have just rediscovered it.

There are 2 options for the fix. Either

1. keep usingarray_view.pyobj() but store the returned object and decrement the reference count myself, or
2. use a new functionarray_view.pyobj_steal() that returns a stolen reference.

I've opted for the second option as that keeps all the explicit array reference counting within thearray_view class.