Performance improvements

Opening an issue to keep a note of performance improvement ideas. Feel free to edit and add to this list:

record cell index and re-use to look up multiple raster values (see benchmark notes below)
push multi-geometry / DataFrame-type operations down to C++ implementation (accept Numpy/GeoSeries/GeometryArray?)

Rough benchmarking for ~570,000 edges, split to ~600,000 splits on ~1km² grid. (OpenStreetMap Bangladesh highways, only trunk to tertiary).

operation	implementation	time
Split to long GeoDataFrame	`snail.multi_intersections.split_linestrings`	1min 13s ± 2.6 s
Look up cell index for splits	`splits.geometry.apply(lambda g: list(get_cell_indices(g, r.width, r.height, list(r.transform))))`	59.8 s ± 166 ms
Get raster values using cell index	`splits.cell_index.apply(lambda i: data[i[1], i[0]])`	364 ms ± 15.1 ms

nismod / snail