Salvage the secondary extraction for the compressed sparse reader somehow

[LTLA]

To stop it from being completely catastrophic as it is now. Suggest just doing something like in that Gist that I have floating around; just create a new CompressedSparseMatrix and fetch rows from it. Every request for something outside the cached range will trigger a full index read, but that's still better than the current state of affairs.

[LTLA]

Should be possible now with oracles to pick, e.g., the next 1000 entries and cache them.