fix handling of 0-label predictions

[sfluegel05]

This came up for the ensemble - if someone selects 0 models, the PredictionSmoother will fail:

chebai/result/analyse_sem.py", line 224, in __call__
    preds = preds_masked_succ.max(dim=2).values
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
IndexError: max(): Expected reduction dim 2 to have non-zero size.

This can be easily avoided by skipping predictions where the label dimension is 0