diagonal disps implementation is suboptimal

[ExpHP]

rsp2/src/tasks/math/displacements.rs

Lines 119 to 127 in ba38408

	// !! FIXME !!
	//
	// Strictly speaking, choices from previous iterations of the `'three` loop could
	// render an arbitrary number of directions from each "star" of directions
	// redundant, requiring us to re-sort the choices to find the new next best choice.
	//
	// It has now been seen that this actually occurs on AB graphene.
	//
	// !! FIXME !!

This theoretical edge case was hit on the very basic example of AB-stacked bilayer graphene, and my code produces 6 disps on both diag and diag-2 where phonopy produces 4.

I'll leave it be for now because it does not affect correctness, only speed, and on BLG the number of displacements saved is O(1) (~4 total regardless of unit cell size). Later, when I start working on more layers, it is possible that structures with an odd number of layers might see savings on the order of O(N). At that point I'll be able to reevaluate whether I want to fix the diag feature, or just rip it out as unnecessary complexity.