add tooling to access and compute on large embeddings

[kevinyamauchi]

In order to support larger-than-memory embeddings, we need some tooling for performant lazy loading. A few features/characteristics:

• should support most numpy style indexing
• should allow for simple queries (e.g., ball point)
• must support zarr both local and remote (e.g., S3)

Libraries for IO:

• TensorStore. this seems like it's the most flexible backend-wise
• xarray
• dask (@kephale noted some performace limitations with large numbers of chunks in the napari tiled rendering work)

[kephale]

For the size of typical tomograms we might be ok with Dask

[kephale]

I'm up for giving TensorStore a whirl.

[kephale]

I have been able to use this album solution to segment whole tomograms:
https://album.cellcanvas.org/cellcanvas/segment-tomogram/0.0.4

after computing the embedding with this album solution:
http://cold-storage.kyleharrington.com/tomotwin/generate-embedding-zarr/0.0.15