The Impact of Off-Policy Training Data on Probe Generalization