Learning from Non-Random Data in Hilbert Spaces: An Optimal Recovery Perspective