A Thread-Register Decoupled GPU Execution Model for Efficient Tensor Computation