Introduce an utility Saver class, that does what tensorflow's Saver
does: keeps track of saved checkpoints in a separate file, and
deletes the old ones before saving a new one.
print() uses stdout by default, which has two problems:
- it is not flushed until later (so messages don't show, or don't show
up in order with other loggers)
- it conflicts with stdin/stdout usage by `decanlp server --stdin`
mehrad