* It seems some other change breaks the async load. It will cause data
  error in GPU mode.

This reverts commit ab279bee.

* Because currently Paddle only use protobuf 3.

* It gets wrong command line arguments before.

* Fix #903

 * gradient_clipping_threshold should be allowed to set with parameter-grain

The chunks of types in excluded_chunk_types will not be counted in
ChunkEvaluator. This is useful for tasks such as SRL, in which chunks of
type V (verb) will not be taken into account in evaluation.

* math => layer_math
* Fix #903

* Partially fix #622

In order to keep consistent with other evaluators, an extra parameter
label is add to chunk_evaluator. Document and demos are modified
accordingly.

These three parameters have already been moved to model_average. Leaving
them here will cause duplicate keys in kwargs