Created by: reyoung

There are several printer evaluators for debugging Paddle network. The GradientPrinter is one of them. It will print input layers' gradient and sequence start position now. I think the GradientPrinter should only print gradient, nothing more. The reasons are:

• The sequence information in GradientPrinter is NOT COMPLETE. If GradientPrinter print sequence positions, how about the nested(sub) sequence positions? Should we add the nested(sub) sequence position?
• If a user wants to print sequence information in the evaluator, we could provide a SequenceInfoPrinter. SequenceInfoPrinter is not provided now, but ValuePrinter print the argument value(a.k.a Neural Network Activations) and sequence information now. Users can use ValuePrinter to print sequence information now.