Improved Prototypical Semi-Supervised Learning with Foundation Models: Prototype Selection, Parametric vMF-SNE Pretraining and Multi-view Pseudolabelling