Remove eos padding when running generation workflow

Created by: suchenzang

We prefix an eos token when running generation (https://github.com/fairinternal/fairseq-py/blob/roller/consolidated_interactive/fairseq/tasks/language_modeling.py#L319). This should be removed.

• Per @stephenroller: It's worth trying but I consider it low priority now.