Puncc

- Add the Least Ambiguous Set-Valued Classifiers (LAC) method in conformal classification module.
- Add theory overview of LAC and Conformal Anomaly Detection (CAD)
- Several documentation improvements

* Conformal object detection tutorial

* Fix issue 57

* Update documentation overall

- Fixed negative sigma issue in `LocallyAdaptiveCP`. The exception `"All MAD predictions should be positive."` should not be raised anymore: the associated negative nonconformity scores are ignored during calibration and result in an infinite prediction interval during inference.
- Added documentation for experimental `PyTorchPredictor`.

- Fixed incorrect pre-trained status check in `DualPredictor` (and `MeanVarPredictor`). Wrapped underlying models must have their `is_trained` status set to True to consider that `DualPredictor` (or `MeanVarPredictor`) is pre-trained.
- Further tests for pre-trained status checks were added.
- Changed `pkgutil.find_loader` with `import lib.util.find_spec` as find_loader is deprecated and will be removed in Python 3.14.

Simplified API procedure for conformalizing pretrained underlying models: instantiating a `ConformalPredictor` object in this situation is performed with a splitter set to `None`.