Make partial observations in the data likelihoods work with the DiagonalEK1

[nathanaelbosch]

This should probably be done after #302: If we handle the solution object better and give it the actual covariance structure it deserves, then implementing this might be much easier.