Refactor the "random" method of Rotations (low-priority)

[pc494]

It currently uses rejection based sampling, I would prefer the method described in:

https://mathworld.wolfram.com/HyperspherePointPicking.html

[hakonanes]

I haven't used Rotation.random(), does it not produce the results you would like, is it inefficient, or both?

orix/orix/quaternion/rotation.py

Lines 531 to 547 in 05311f6

	def random(cls, shape=(1,)):
	"""Uniformly distributed rotations.
	Parameters
	----------
	shape : int or tuple of int, optional
	The shape of the required object.
	"""
	shape = (shape,) if isinstance(shape, int) else shape
	n = int(np.prod(shape))
	rotations = []
	while len(rotations) < n:
	r = np.random.uniform(-1, 1, (3 * n, cls.dim))
	r2 = np.sum(np.square(r), axis=1)
	r = r[np.logical_and(1e-9 ** 2 < r2, r2 <= 1)]
	rotations += list(r)
	return cls(np.array(rotations[:n])).reshape(*shape)

I'm not entierly sure how to implement the equations in the link, but I guess you do. The hypersphere point picking should be added as an alternative method to the existing one, I assume.

[pc494]

The current version is correct, it's doing the Marsaglia (1972) method from:

https://mathworld.wolfram.com/SpherePointPicking.html

which gives the correct answer. However it's slower than the version I'll replace it with because it involves discarding entries (also it's a while loop).

Consider this nothing more than house keeping that came up while I was working with some of these bits of the code. I'll implement the fix in my big bundle of things.