This adds url, data and parser options to the source, and makes
EPSG:4326 the default projection. It also adds a prepareFeatures
method, which is used to load/parse data once the target
projection is known.
By maintaining separate R-Trees for each geometry type, we cover
the most critical query from the canvas renderer at higher
speeds. Using the filter's own evaluate method, we can now also
serve arbitrary queries, but they won't be fast.
Still needs unit tests.
Also trying to reference features directly here instead of
storing feature ids in the index. We'll see if this is good when
we have an index for tile coords as well and want to filter
with multiple filters, accessing multiple indices.
This needs more work still - see inline TODOs. It also has a
major bug - rendered vector features do not scale and do not
change their offset during panning. So only the initial view is
correct.
