Use tf.clip_by_value again in PR curves demo

[chihuahua]

Previously, #1132 had circumvented the use of tf.clip_by_value within the PR curves demo because it was buggy. However, that behavior seems to have been fixed. See tensorflow/tensorflow#18527.

Hence, we make the demo use tf.clip_by_value once again.

[nfelt]

LGTM to revert back to using clip_by_value - and actually, I investigated a bit and I think the underlying problem is specific to the PR curve demo/test, versus a problem with clip_by_value per se.

The root of the issue is that tf.set_random_seed() only makes random values (as produced by functions like tf.distributions.Normal.sample) deterministic within the context of the exact same graph structure. If the graph structure changes, specifically if nodes are added/removed such that the node ID of the random op changes, then this directly changes the op-specific "deterministic" seed which gets used when the overall graph seed is fixed: https://github.com/tensorflow/tensorflow/blob/v1.8.0-rc1/tensorflow/python/framework/random_seed.py#L68

Our PR curves plugin integration test expects the values to be 100% deterministic, which means that graph structure changes that change node ID numbering of the sample() ops will break the test. I believe the changes in tf-nightly to clip_by_value() triggered this because clip_by_value() changed in terms of its internal op structure and that changed node ID numbers - but it's not the "fault" of clip_by_value().

In the short term, I figured out how to extract the op-specific seed values and will send a PR that uses fixed op-specific seeding so that adding/removing nodes won't break the test.

However, I don't think even with fixed graph and op seeds TF actually guarantees random op output values (I believe it guarantees only that they are deterministic for a given TF version, not that they won't change across versions). So the preferred fix would be removing the reliance of the test on randomness-based data generation, and instead do something like use a fake multiplexer populated with hardcoded fake PR curve tensor summaries.

[chihuahua]

Thanks for investigating! Running the test on some static data seems like a good idea then.