`common_chunks` on `SubDiskArray` can be a large fraction of io cost

[rafaqz]

This is where the left hand side of the broadcast is a Matrix and the right is a SubDiskArray

Probably common_chunks should never be called for this case?

(for context this is completely killing the performance reading thousands of 500*500 windows from a large lazy disk array into a preallocated matrix)

[rafaqz]

Additionally the same common_chunks call is being made at multiple levels

[rafaqz]

I guess the problem is TIFF defaulting to 1 chunk per column, so there are very many chunks covered by a small window.

julia> DiskArrays.eachchunk(rast.qualities)
1×31824 DiskArrays.GridChunks{2, Tuple{DiskArrays.RegularChunks, DiskArrays.RegularChunks}}:
 (1:25811, 1:1)  (1:25811, 2:2)  (1:25811, 3:3)  (1:25811, 4:4)  (1:25811, 5:5)  …  (1:25811, 31822:31822)  (1:25811, 31823:31823)  (1:25811, 31824:31824)