compute_stars: input deperms violate the group axioms

[ExpHP]

Pulling this remark out from here: #64 (comment)

To the best of my knowledge at the time I ran into this, I figured this should be completely impossible; compute_deperms should not be capable of returning inconsistent permutations (IIRC the current implementation is even paranoid about supercells). But it happened, so I must be mistaken about something! 😛

(Unless... maybe it could happen if two atoms somehow ended up right on top of each other? I'm not sure what it would do in that case. I'd have to look over the code again)

[ExpHP]

Turns out the structure in question was using the conventional cell of a BCC solid. rsp2 is currently only designed to support primitive cells as input, so the main issue was that the sparse diagonalizer was not bailing out on supercells like the phonopy diagonalizer does.