CARL: Camera-Agnostic Representation Learning for Spectral Image Analysis