Neural Transducer Training: Reduced Memory Consumption with Sample-wise Computation