Stabilize the data unit tests

[jkamalu]

This PR makes a few changes to address potential sources of instability/error in the data unit tests.

1. The unit tests touched in this PR only need to be run on a single process, so in the multi-rank setting, we only do the work on one rank and return early on all others
2. We revert to making an attempt to find test tokenizer vocab and merge files locally before downloading them from the web
3. We fix the local path for the BERT vocab file
4. We standardize somewhat the object storage client spoofing

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[jkamalu]

/ok to test e485838