distributed rtree

[keewis]

• towards distributed rtree #18

This contains a draft of the distributed rtree. Main remaining issues is constructing the indexes for the chunks on a worker, and assembling the final grids as a sparse matrix (though I guess the latter may be borrowed from VirtualiZarr?)

[keewis]

the distributed tree works already, but I'm sure the code – especially the part using dask – could be further optimized. I still have to check whether the results actually make sense (see also #7).

Additionally, right now the chunk boundaries are computed as the union of the grid geometries, which takes quite a while for large chunks. It might thus make sense to allow passing pre-computed chunk boundary polygons, which would make the index truly lazy.

cc @rsignell, @maxrjones, @norlandrhagen, @TomNicholas, in case you're interested

[keewis]

single-chunk as well as in-memory indexing appear to return the right results, but multi-chunk does not quite work yet. I suspect the issue is with the chunking; I'm flattening rectangular chunks and assemble the result as one 1D array per target chunk, so that could explain the wrong results.