Optimizers

Compare SGD, momentum, Adam, and AdamW, including optimizer state, learning-rate schedules, clipping, and weight decay.