Using a representation of a low-rank symmetric matrix

[sethaxen]

Currently the C matrix is allocated as dense and updated in-place using rank-1 updates. This is fine when the data is not large (see #8) or when the rank is close to the size of the data, but otherwise it will be inefficient in terms of both storage and computation.

A better alternative would be to have an AbstractMatrix type that stores the rank-1 updates as either one long vector or a vector of vectors and then implement the necessary methods for this to act on other vectors or matrices.

[SebastianCallh]

You are absolutely right and that's a good suggestion. I have not pursued any optimizations yet since I don't fully trust the current implementation (#5).