Evaluating Computational Pathology Foundation Models for Prostate Cancer Grading under Distribution Shifts