[feat] Add a map-based dataloader option

[blefaudeux]

Right now it's only iterable, which limits a lot of the usefulness. We could cache the pages in the process then expose an index driven access (once the initial page walk is done)