Fixed bug in ModelWeightAveraging class that led to corrupted model when metric to watch was NaN/Inf

[BloodAxe]

A class was not checking whether observed metric is a finite number and included a state dict of the model that had inf/nan metric value into averaging.

The fix was three-liner:

val = float(validation_results_dict[self.metric_to_watch])

if not np.isfinite(val):
  return False, None

[Louis-Dupont]

LGTM, just make sure the greater_is_better without default is fine for classification (as raised by ofri)