Parameter-Efficient Self-Supervised Adaptation for EEG-FM under Fixed Computational Budgets