Add warning message for run_qa.py

[jla524]

What does this PR do?

Fixes #29801 (issue)

When running with --load_best_model_at_end, --metric_for_best_model is set to "loss" by default. However, loss is not supported by squad or squad_v2.

We can throw a warning to let the users know to explicitly set the metric to one supported by squad or squad_v2.

Who can review?

@ftesser and @ArthurZucker

[SunMarc]

Make sense, LGTM ! Instead of raising an error, let's raise a warning if we don't put all available metrics (e.g. squad2 have other metrics such as HasAns_exact,....)

[jla524]

Thanks @SunMarc! I've updated the PR to raise a warning instead. It'll show a warning like this:

/Users/jacky/repos/transformers/examples/pytorch/question-answering/run_qa.py:636: UserWarning: --metric_for_best_model should be set to one of ('exact_match', 'f1')
  warnings.warn(f"--metric_for_best_model should be set to one of {accepted_best_metrics}")

[ArthurZucker]

Thanks!