Fix memory leak in ManagedRaster

[emlys]

Fixes #483
This fixes a couple of memory leaks:

In the next method of the NeighborIterator classes, we allocate memory for a new NeighborTuple to store the current value, and delete the old NeighborTuple. If a NeighborIterator wasn't iterated all the way to the end, the final NeighborTuple would never get deleted. I added a destructor to NeighborIterator that deletes the final NeighborTuple if necessary.

Also, the geotransform attribute was not being freed. This was pretty minor since there's only one of them per ManagedRaster instance.

I tested this by building natcap.invest against this branch of pygeoprocessing, running SDR on the sample data with valgrind, and observing that there are no longer any loss records that mention ManagedRaster.h.