…n feed_dict or 1 epoch using readers.

* Improving PrintTensor monitor, to support tags for printed tensor (e.g. passing {'loss': loss_op} will now display loss = %f instead of full name of the op).

* Improving ValidationMonitor to support various metrics, minimization/maximization, naming to run multiple validations.

* Make sure early_stopping test actually early stops. Updated example as well.
Note, test is unstable, so exact number of steps it stops are non reproducible. See stability tests for more examples of issues.

* Added GraphDump monitor for in-depth debugging.

* Added stability test to make sure the same model trained on the same data given exactly the same results. Note: it's all super unstable, increased tolerance to just make it pass. Possibly issues with numerical stability in TF.

* Changed max_steps in graph_actions.train into steps, which adds that many steps to the training (instead of just defining max steps). For Estimator this returns previous logic, where fit(..., steps=100) followed by fit(..., steps=100) will result in 200 steps trained.
Change: 124197624