Fix evaluation issues

[Tixxx]

Description:
Fix evaluation issues

Motivation and Context
Override drop ratio in eval with 0 so that Dropout behaves as an Identity node so that evaluation results are correct.
Added unit tests for python frontend.

[SherlockNoMad]

Overall, I find this approach a bit too intrusive to the framework. It's injecting a kernel level control from the RunOption. I think this is breaking the original design principle that an op is self-descriptive and stateless. Input+Attribute alone should determine the behavior of an op.

If this flag is introduced, the op's behavior would depend on the config set by run time.

For dropout, we have a "ratio" input to determine if it should invoke training mode.
Similarly for BatchNorm-12, we introduced another input "is_training_mode".

To make these nodes to operate in the eval mode, we can override the inputs of there nodes. Initializer of a node can still be override with graph feeds.

[jessebenson]

I agree with Sherlock, and he summarized the design quite well.

There's a relatively small number of operators that need different behaviors in training vs inferencing/evaluation mode. We should favor controlling those behaviors with operator inputs, and if we need to control it per session.run(), then we can add that as a graph input. As Sherlock said, we can have an initializer (for the default value) and still add it as a graph input so we can override it with the input feeds.

[liqunfu]

Do not forget to rebase and target to the master branch

The base branch was changed.