Consider different architectures for greater performance

Currently this iterates over tiles (in parallel), uses an rtree to find all matching features, then builds each tile.

What about:

- going through each feature once and deciding all the tiles it belongs to? (from the bbox, can figure that out for any zoom level cheaply)
- iterating over the rtree structure directly (kind of like https://github.com/acteng/will-it-fit/blob/d91dbf0d70828d4b92a0de5f26a28bf274f89540/data_prep/dissolver/src/main.rs#L166)
- follow the [planetiler architecture](https://github.com/onthegomap/planetiler/blob/main/ARCHITECTURE.md) (which has the best performance today for huge inputs, right?): loop over each feature, loop over each zoom level, writes encoded per-tile intermediate output somewhere. Then the sorting + feature dropping + compression + writing happens in a later pass.