This replaces the previous primitive ad hoc hacks we needed to get started.
In particular this can parse the full set of level-related OSM tags
including labels, ranges and repetition. This also can now properly handle
intermediate levels (e.g. half-way levels in stair cases), and sort the
level stack properly. We also explicitly filter node-only levels where
we cannot render anything meaningful.

This still needs proper API to access the level information for the UI
though.