Create distributed data filtering on JZ

[huu4ontocord]

Assuming we will be doing large amounts of data filtering dynamically on JZ, we may need to do this we will need to adapt this:

bigscience-workshop/Megatron-DeepSpeed#60

To run the filtering pipeline.

[huu4ontocord]

First start with this issue: #31
Then move to make multi-node under this issue.

[huu4ontocord]

@adammoody any interest in helping with this issue? We actually need to do data dedup, and need to do it in a distributed way.