Get Strong Typed Metric From paddle::Evaluator

[reyoung]

In current Paddle API in Python, we should get the strong-typed metric when training, such as Error Rate in float.

We should add some interfaces to C++ class paddle::Evaluator to get the metric result firstly, and then we could expose them to SWIG/Python.

Each current evaluator could contain zero to many metrics inside.

• each evaluator should include one metric result, like classification evaluator, contains ErrorRate.
• There are some evaluators just used for debugging, they often called XXXPrinter.
• There are some evaluators will contain many metric results, like precision-recall evaluator. It because some metrics are calculated together.

So basically, the metrics in Evaluator should be a map or dictionary, the key is a metric name, the value is the strong-typed result.