beam search should support skip_values

[szha]

When decoding, a common use case is to skip some of the output values, such as padding, end-of-sequence, unknown token, etc.

[szha]

One possible solution is to manually set all the scores corresponding to the indices in skip_values to be negative numbers of large absolute values, so that the corresponding beams don't make the cut.

[szha]

https://github.com/dmlc/gluon-nlp/blob/master/gluonnlp/model/beam_search.py#L240-L252