Remove sigmoid in BinaryPrecisionRecallCurve

[PussyCat0700]

What does this PR do?

This PR addresses #3179 by making the use of sigmoid normalization optional in BinaryPrecisionRecallCurve. Specifically, it introduces a normalization parameter to _binary_precision_recall_curve_format and disables automatic sigmoid application in BinaryPrecisionRecallCurve.

Changes

• Adds a normalization: Optional[str] = "sigmoid" argument to _binary_precision_recall_curve_format, allowing users to override the default behavior.

• Passes normalization=None in BinaryPrecisionRecallCurve to preserve raw logits and prevent batch-wise distortion during .update() calls.

Notes

• This fix only applies to BinaryPrecisionRecallCurve, which is the class I directly encountered issues with.

• I’ve noticed that other metric classes may also apply sigmoid implicitly in ways that could cause similar inconsistencies across batch-wise updates, but since I have not tested those cases, I am leaving them unchanged to avoid unintended side effects.

---

📚 Documentation preview 📚: https://torchmetrics--3182.org.readthedocs.build/en/3182/

[PussyCat0700]

python -m pytest -v --color=no tests/unittests/classification/test_roc.py::TestBinaryROC::test_binary_roc>
Seems that a few unit tests weren't passed because the test cases have only considered the results after sigmoid to be correct. So it is impossible to pass the unittests currently, I suppose?

[Borda]

pls add entry to changelog

[PussyCat0700]

pls add entry to changelog

Done