Accelerate ND-Parallel: A Guide to Efficient Multi-GPU Training