fix ConfusionMatrix and StatScores for num_classes > 16

[vincentvaroquauxads]

e.g. if preds or target is uint8 and num_classes > 16, unique_mapping overflows

What does this PR do?

Fixes #1510

I would suggest to define :
unittests.helpers.NUM_CLASSES=17
and add something like the following to unittests.classification.inputs._multiclass_cases

    pytest.param(
        Input(
            preds=_logsoftmax(torch.rand(NUM_BATCHES, BATCH_SIZE, NUM_CLASSES), -1),
            target=torch.randint(high=NUM_CLASSES, size=(NUM_BATCHES, BATCH_SIZE), dtype=torch.uint8),
        ),
        id="input[single dim uint8-logits]",
    ),

But it breaks a lot of tests...

[justusschock]

@vincentvaroquauxads thanks for checking it out and submitting this PR. Would you be interested in enabling the change with num_classes=17 and fixing the issues arising from this? It should almost always be dtype issues.

[codecov]

Codecov Report

Merging #1521 (0803189) into master (a9def1c) will decrease coverage by 36%.
The diff coverage is 100%.

Additional details and impacted files

@@           Coverage Diff            @@
##           master   #1521     +/-   ##
========================================
- Coverage      87%     51%    -36%     
========================================
  Files         216     216             
  Lines       11334   11334             
========================================
- Hits         9845    5752   -4093     
- Misses       1489    5582   +4093     

[vincentvaroquauxads]

I investigated a little bit more and some tests fail because ignore_index=-1 and target.dtype=uint8.
I changed target.dtype to torch.int8, but I still have issues...

Actually, after hours, I finally realized that, using the original master branch, tests are broken in my environment.
It might be linked with this warning:
sklearn/metrics/_classification.py:1318: UndefinedMetricWarning: Recall is ill-defined and being set to 0.0 in labels with no true samples. Use zero_division parameter to control this behavior. (scikit-learn 1.0.2)

Sorry, no time to investigate further... And it's another issue (no link with this PR)...

I pushed the unit tests, if anyone can try them on a clean environment.