Add ModalDownsampler

[superdump]

Something like this I think. In the test I was trying to use 1x1x1 lod0 chunks and a 2x2x2 superchunk and lod0 extent, thinking that it would downsample the 2x2x2 extent of lod0 to a 1x1x1 lod1, but I hit the assertion on the dst shape being Point3i::ZERO. I don't understand why.

[superdump]

I've tried it for my blocky voxel terrain stuff and it's really slow. Perhaps it's useful for someone?

[bonsairobo]

Maybe you could add a benchmark in building_blocks_storage/benches/sampling.rs so we can compare actual numbers?

I also suspect that using the SmallKeyHashMap would help. It would be even better if we could assume that the voxel type was an index into an array.