Add a utility for performing randomized rotations of vectors.

[vekterli]

@havardpe and/or @arnej27959 please review.

This aims to implement the "exercise left for the reader"/"draw the rest of the owl" $y \gets \Pi x$ and $x \gets \Pi^\top y$ parts found in most quantization papers (where $\Pi$ is a random rotation matrix, $\Pi^\top$ its inverse/transpose, and $x$, $y$ are pre/post rotation vectors, respectively).

We use the fast Walsh-Hadamard transform (FWHT) as the work horse algorithm to implement rotations. The FWHT is a deterministic transform and not a "true" matrix multiplication, but we can get results as if we had used a random matrix multiplication if we pseudo-randomly flip the signs of the input vector prior to rotation.

Fast Hadamard transforms are only defined for power-of-two vectors, so for vectors where this property does not hold we use multiple overlapping sub-transforms to achieve a statistically similar result. Note that this requires more computations (up to 2x) than if the vectors were powers of two.

The partially overlapping transforms are inspired by the RaBitQ rotation algorithm, but we use a different strategy for "bridging the gap" (well, technically bridging the overlap) across the sub-transforms; see the comment for the Rotator class for details.

The quality of the rotator algorithm has been estimated by sampling the expected Gaussian vs actual post- rotation coordinate distributions for several scenarios (sparse vectors, shifted mean normal distribution), and then computing the Kullback-Leibler divergence between the two distributions. The intuition and assumption™ being that a divergence close to zero means that the rotation has the desired properties for our quantization purposes.

This implementation is not yet optimized for speed, and we may change things around based on further experiments. But it should be a functionally complete starting point.