Migrate asr bleu

[CalderJohnson]

Why ?

As part of seamless_communication, I created a script that evaluates the model's translation accuracy using whisper ASR and BLEU metric.
https://github.com/CalderJohnson/seamless_communication/tree/implementing_eval_script/scripts/m4t/evaluate
In order to take advantage of powerful stopes features such as caching, the launcher, and parallelism, I've implemented a stopes module that can run this evaluation protocol on any number of desired language directions. This will aid in research with the model as it will facilitate quick, parallelized measurement of translation accuracy.

How ?

I've created a simple stopes pipeline consisting of a single module that runs the compute_asr_bleu script on the specified language directions in the fleurs dataset.

Test plan

How did you test your changes ?

I ran tests on a few language directions, ensuring the output was as expected. Reproduce with your own testing config file such as:

# @package _global_

datasets:
  test1:
    lang_dir: "fra-ita"
    num_data_pairs: 2
  #any number of others...

place it in a directory inside of conf/ eg. "testing"

and run python stopes/pipelines/asr_bleu/asr_bleu.py +testing=testfile +output_dir=.`