Uncovering Neural Scaling Laws in Molecular Representation Learning