Multi-View Latent Action Pretraining for Cross-Embodiment Robot Learning from Videos