Identify which shard data belongs to

**Is your feature request related to a problem? Please describe.**

I'm training on a large dataset made of multiple sub-datasets.
During training I can observe some jumps in loss which may correspond to different shards.


My suspicion is that either:
* some of the sub-datasets are harder for the model than others
* some of the sub-datasets are not formatted properly

I'd like to identify which shards correspond to those jumps.

**Describe the solution you'd like**

It would be nice to have a key associated to each data sample or data batch containing details on where the data comes from (shard idx + item idx within the shard).
This should be supported both in local and streaming mode.

**Describe alternatives you've considered**
A fix would be for me to add myself details (shard id, sample id) as part of each data sample.
The inconvenient is that it requires users to process/reupload every dataset when they need this feature.