Starting from complete base layer, we want to move up the pyramid, producing
zoom-appropriate vector tiles at each level.  There is a fair amount of data
duplication at low levels, since each key that intersects a geometry will have
a separate copy of the geometry.

The general action of moving up the tree is to re-key according to the next
coarsest key layout (divide key col/row by 2), and to drop duplicate
geometries, presumably by the (node/way/relation) id that generated it.
Following this step, a filter can be applied to drop geometries that will not
be used at higher zoom levels.

There is also a potential for simplifying geometries at the same time as the
filter operation.  It's not clear what the simplification criteria should be
at this point, but some-scale aware criteria (tuned to the vector tile pixel
grid) would be ideal.

A translation from geometries to vector tile should also happen at each zoom
level.  This will involve cutting geometries to the desired region—either a
buffered zone, or to the edge of the VT boundary.  A selection criteria will
also be supplied to determine the entities that exist at each zoom level
(buildings at zoom 15 and finer, highways only at high zoom levels, dirt roads
at zoom 14 and finer, etc).

Schematically:

```
       ┌──────────┐
       │ Zoom z-1 │
       └──────────┘
            ↑
      simplification
            ↑
        gathering
            ↑
        filtering
            ↑
       ╔═══════════╗                          ┌─────────────┐
       ║  Zoom z   ║ → selecting → clipping → │ VECTOR TILE │
       ╚═══════════╝                          └─────────────┘
```