fix(drift): make ABCD usable without PyTorch, and default to PCA

[hmgomes]

ABCD reconstructs instances with one of three encoder-decoder models, but only the autoencoder needs PyTorch. PCAModel and KernelPCAModel use scikit-learn, which is a core dependency. All three shared a module whose top-level import torch.nn put the whole detector behind the torch gate, so a default install lost CapyMOA's only multivariate drift detector even for configurations that never touch PyTorch.

AutoEncoder moves to its own module. It subclasses nn.Module, so the import cannot be deferred into a method; a separate file is what lets abcd.py load it lazily and only when model_id="ae" is requested. Asking for that model without the extra raises an error naming the model rather than the detector:

OptionalDependencyError: PyTorch is required for ABCD's autoencoder model (model_id='ae').
Install it with: pip install capymoa[torch]

The default changes from "ae" to "pca", so a plain ABCD() works in a default install. That alters which model an existing caller gets and belongs in the release notes. On the univariate stream in the drift detection notebook the two behave identically (1 detection, 2 warnings, first at 1011), so its stored output is unchanged.

Verified in a real torch-free virtualenv, not only with the sys.meta_path simulation: ABCD imports, stays in __all__, and pca/kpca/default all run with torch never entering sys.modules.

Closes the follow-up left open by #370, which supplies capymoa._optional and OptionalDependencyError.