Better restore and checkpoint behavior

• Fix one-off error where checkpoint saved a step early.
• Fix issue where restore_from='fresh uses the counter from a previously-trained checkpoint.
• If restore_from='latest , steps will now train for the specified amount of steps, instead of the training until the specified number of steps. (#13, #14)