Check on validity of using bbox as a fallback to proj:bbox

[TomAugspurger]

Right now the computation of the output size uses the bboxes from the individual stac items. In the best case scenario, the items have proj:bbox and we use that. But since proj:bbox isn't required, we fall back to the item-level bbox and then reproject it according to the crs. In some testing on NAIP data in Azure, the reprojected bbox isn't really close to the proj:bbox. The 'reproject bbox' route is also a lot slower, though I haven't profiled things.

Maybe we just require proj:bbox?