add FreDF loss combining time-domain MSE and frequency-domain MAE

[DresdenGman]

Following up on #1301 — implements the FreDF loss proposed by Lin et al. (2023).

FreDF combines time-domain MSE with frequency-domain MAE via torch.fft.rfft, weighted by a tunable alpha parameter (default 0.5). When alpha=0 it reduces to pure MSE; when alpha=1 it reduces to pure frequency-domain MAE.

Added FreDF to all in losses/pytorch.py, inheriting BasePointLoss to stay consistent with the existing loss API.

Closes #1301

[CLAassistant]

All committers have signed the CLA.

[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]

@DresdenGman thanks for the contribution! Could you add a few tests in neuralforecast/tests/test_losses/test_pytorch.py please. Especially for when alpha is 0 and 1, that will validate the implementation is sound. Thanks!

[DresdenGman]

Added three tests in test_pytorch.py: alpha=0 (reduces to MSE), alpha=1 (reduces to frequency-domain MAE), and perfect forecast gives zero loss for any alpha.

[marcopeix]

@DresdenGman, see my comment. Seems like you're adding an entirely new test file, but you should modify the existing test file.

[marcopeix]

The tests are good, but it seems like the entire file is new. That's not right. You should modify the existing file: https://github.com/Nixtla/neuralforecast/blob/main/tests/test_losses/test_pytorch.py