Scaling laws for language encoding models in fMRI