Add pre_scan routine to SequenceScanner.

support multi-field for inference

stride pooling for seqlastin and seqfirstin.

Add list type of feeding

add wmt14 trg_dict

nce_layer cannot accept neg_distribution

reading source code with woboq

fix PadOp bug on Gpu

It is no need to config num_classes in hsigmoid

It should be always equals to label's size.

Add cross-compiling support for arm architecture.

support Beam search in v2 api

Remove unecessary code to generate freq_dict.

Speed up int scanner of data converter.

* Pre-allocated buffer for int scanner

Using linb::any/std::any instead of FunctionConfig

Refine NCE Layer

* Also test std::vector
* Use std::vector to PadConf