Diffusion-based dimensionality estimation, to support graphs

[professorwug]

We currently use the kNN dimensionality estimator from skdim - but this has problems:

• It needs an array of points
• It laboriously reconstructs the graph from those points (duplicating our own graph)
• It uses shortest path distances, when we have better geodesic estimates

There are two ways of resolving this:

1. In the short term, use the same algorithm, but use the precomputed geodesic distances. This would solve all of the above issues.
2. In the long term, find something more robust than kNN dimensionality estimation.