[nvictus]

Restores tile batching optimization as found in the legacy cooler tileset, but improved and with greater flexibility.

Tiles from a request batch are grouped by zoom and transform. Then, within a group the reader merges queries whose bin ranges are contiguous in the cooler's own bin ordering. Finally, the merged rectangle is sliced back into per-tile blocks.

Consulting the file's internal ordering rather than what we project on the canvas makes it safe for us to project a chromosome order that differs from the file's (though, this option is not exposed yet!) This is now exposed via an optional chromsizes argument to the tileset constructor. Tiles that are adjacent on the canvas are then not adjacent on disk, and simply would not merge.

We re-use cooler's internal query engine machinery for the design:

1. Because pixels are stored CSR, merging fetches of horizontally-adjacent bboxes eliminates multiple fetches of the same matrix rows from disk. For bboxes crossing the diagonal in symmetric-upper mode, vertical consolidation also eliminates redundant I/O.

2. As an additional optimization, fetches go through cooler's internal query engines rather than Cooler.matrix() to read the internal index only once for the whole batch.

The original, and simpler, non-batched query mode is preserved by passing batched=False.