Allowing extract_array calls to use pre-indexed grid information?

Start by reading the extract_array and rowAutoGrid entry points, then review the approach described in LTLA/beachmat#20 for pre-indexing dgCMatrix row blocks. A complete change would define a general way for suitable DelayedArray backends to use pre-indexed grid information and demonstrate improved row-based extraction performance.

The recent conversation in theislab/zellkonverter#34 reminded me of some work I did in LTLA/beachmat#20. Briefly, the idea was to speed up row-based block processing of dgCMatrix by performing a single pass over the non-zero elements beforehand to identify the start and end of each row block in each column. This avoids the need for costly per-column binary searches when each row block is extracted in the usual way, and gives a ~10-fold speed-up in row-based processing of dgCMatrixes.

Now I'm wondering whether this approach can be generalized somehow so that other DelayedArray backends can benefit. Perhaps functions like rowAutoGrid() can decorate the grid object with extra information that allows extract_array to efficiently obtain the necessary bits and pieces, if a suitable object like a dgCMatrix is passed?

Happy to give this - or other ideas - a crack with a PR if there is some interest.