I had proposed doing something like this in #5185 (comment), but @tacaswell thought that didn't seem quite right. I guess I understand the objection from a purity point of view, but on the other hand most of the C++ code assumes a fix size for all the dimensions past the first one (i.e. an array of bounding boxes is always Nx2x2, so if it's passed an array of 1x0x1, I'm fine with it behaving as if it's 0x2x2). And handling the special case of "a zero in any dimension" in one place (the size method) as it's done here is much better than handling it everywhere that uses array_view.

For completeness, we should still do something like half of #5246, which also fixes this bug but checks the expected lower dimensions. But with this, we wouldn't need to revert the ensure_3d -> atleast_3d change. I'm going to create a new PR that combines this one with that one so it can all be tested together.

Closing in favor of #5274.