New indexing

[vince62s]

This PR has the following purposes:

• get rid of the old batch["indices"] key which was not actually used.
• introduce the following keys
  batch["cid"]: identify the corpus id of the example
  batch["cid_line_number"]: identify the corpus id line number
  these two keys will be helpful to retrieve the lines where a training stopped and resume training at those lines
  one other key:
  batch["ind_in_bucket"] de facto replaces the "indices" but represents the index in the current bucket

This PR hence will sort buckets ALSO at inference (was doing it at training) which will speed up the inference and then after translation we reorder the translations from these "ind_in_bucket".