The Golden Subspace: Where Efficiency Meets Generalization in Continual Test-Time Adaptation