fixes for validation engine and for using torchrun

[jgmf-amazon]

Issue #21

Description of changes:
There are two fixes given:

1. Code was added to the eval_preds function in training_utils.py per the suggestion of @bozheng-hit .With the current main branch, the validation engine is not functioning correctly, because subwords after the first subword are being handled by the convert_to_bio function. Instead, we want to merge/ignore the subsequent subwords and use only the prediction from the first subword. This change sets subsequent subwords to -100.
2. The train, test, and HPO scripts were modified to accept environmental variables for local_rank, allowing them to support torchrun.

I tested the changes with the xlmr_base example configs, both with train.py and test.py.

Once these changes are approved, I'll update the README. Thanks.