Enabling Energy-Efficient DNN Training on Hybrid GPU-FPGA Accelerators