Currently, the --train_from option does not include means of restoring corpora states, hence training resumes from the beginning of the bitexts. This entails resumed models are training on a subset of the available data, unless some manual shuffling is done between each resumption.

Fixes (partially implemented on V2, but never ported here

• add a line index as files are read, pass it along when collating batches, and skip up til this line index upon training resumption
• a more complex refactoring is needed to save the dataloader state, as it would involve communicating all examples in the reservoir — which would be much more costly to communicate.

CC @jrvc