Enhancing Low-Precision Sampling via Stochastic Gradient Hamiltonian Monte Carlo