Created by: punitkoura
Patch Description Adding support for subsharding training data as described in #166 (closed)
Testing steps Added unit tests for jsonl_dataset.py in test_jsonl_dataset.py
Confirmed that the baselines run works as expected.
--UPDATE--
As per discussion in https://github.com/facebookresearch/metaseq/issues/166 , we now ensure that one shard is exhausted before moving on to the next.