Defensive Programming of Model.save()

[colah]

Model.save() is the new recommended way for people to import models into Lucid. While pairing with @Newmu the other day, we ran into a number of issues where we made silly mistakes, but it took a bit of time to debug.

Since Model.save() is such an essential user facing function, I think it warrants heavy defensive programming, good error messages, and documentation to protect against this.

Some checks that would have been nice:

• Check for ":0" in names -- A ":" denotes a TensorFlow tensor, but we actually want ops.
• Check if default session is None (more info).
• Check if graph is empty