Multi component spectral layout (Laplacian Eigenmaps)

We need another primitive to perform the spectral embedding on connected components of the connectivity graph in UMAP. This will help us bring UMAP closer to the reference implementation.

The current spectral clustering implementation computes the eigenvalues on a laplacian of connectivities which originate from a knn graph, but this is not sensitive to datasets that might have various connected components. We need to implement a first-class laplacian eigenmaps implementation that performs an embedding/clustering of all connected components separately.