fix: preserve infinite losses in weighted reductions

[kiwoongyoon]

Title: fix: preserve infinite losses in weighted reductions

Summary

• Preserve signed infinities in _divide_no_nan instead of replacing them with zero.
• Ignore zero-weight entries before multiplying weighted losses, preventing masked infinities from producing NaN.
• Cover unmasked and masked positive/negative infinite predictions with an MAE regression test.

Fixes #1601

Test evidence

• pytest tests/test_losses/test_pytorch.py::test_MAE_infinite_predictions -q — 1 passed
• pytest tests/test_losses/test_pytorch.py -q — 16 passed
• ruff check neuralforecast/losses/pytorch.py tests/test_losses/test_pytorch.py — passed
• git diff --check — passed

Risks or notes for maintainers

_weighted_mean is shared by point losses, so zero-weight non-finite values are now ignored consistently wherever masks or zero horizon weights are used. No public API or dependency changes are included.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[marcopeix]

@kiwoongyoon Thanks for raising the issue and drafting a PR. I think I have a more complete fix in #1603

[marcopeix]

Closing since merging #1603