Spatial-Temporal Masked Autoencoder for Multi-Device Wearable Human Activity Recognition