Towards Precise 3D Human Pose Estimation with Multi-Perspective Spatial-Temporal Relational Transformers