Replace spatial-graph with rstar-python

[JoOkuma]

Depends on kephale/rstar-python#4.

Replaces spatial-graph with rstar-python point and bounding-box indexes while preserving inclusive ROI queries, dynamic updates, exact node deletion, and one-dimensional filters. rstar-python is now a core dependency; the spatial extra is removed, and uv is pinned to the upstream PR commit until 0.2.0 is released. ASV installs both backends so it can compare the spatial-graph baseline with this branch.

Tests: uv run pytest src/tracksdata -q (1293 passed, 12 skipped, 1 xfailed); Ruff check/format pass; targeted asv continuous upstream/main HEAD --quick passes.

[JoOkuma]

Hi @TeunHuijben and @cmalinmayor , any thoughts?
Kyle suggested using it.

[cmalinmayor]

I'm in favor of not requiring spatial-graph as long as it it needs to run C++ at build time, since our Windows users are very unhappy. Presumably the benchmarks show similar speed?

[TeunHuijben]

@cmalinmayor, @JoOkuma,

I created a PR on funtracks that uses this rstar-python version of tracksdata, and benchmarks show that it is faster across all three platforms. Not sure if the funtracks benchmarks cover everything (might have to go to motile_tracker), but is a good sign!

[cmalinmayor]

It really should only affect the GraphArrayView 🤔 Which I'm not sure we benchmark in funtracks at all. The fact that it's faster in funtracks implies we still have some c-compilation time in our benchmarks.
Update: actually looked... Seems like we probably do construct a GraphArrayView since all the faster things have to do with segmentation! Seems plausible!

[TeunHuijben]

Hmm interesting, I'll make a motile_tracker draft PR as well

[JoOkuma]

I did see a few speed-ups compared to spatial-graph.
Thank you both for the input. I'll merge this.