Fix logger verbose level not restored after unpickling

[JayeshSuryavanshi]

Problem

The logger verbose level is not restored after a HierarchicalClassifier is pickled and unpickled (issue #146).

_create_logger obtains the logger via logging.getLogger(self.classifier_abbreviation) (e.g. "LCPN"), which is a process-global singleton, and Python pickles a logging.Logger by name only (restoring it via getLogger(name), discarding the level). So on unpickle, model.verbose survives (a plain attribute) but model.logger_.getEffectiveLevel() does not, it inherits whatever level the current process's singleton happens to have. Because every instance of a classifier shares one named logger, any other instance or process that touched it determines the restored level.

Fix

Add __setstate__ to HierarchicalClassifier that recreates the logger from the restored self.verbose:

def __setstate__(self, state):
    self.__dict__.update(state)
    if "logger_" in state:
        self._create_logger()

The if "logger_" in state guard means only fitted models rebuild the logger. An unfitted model has no logger_, and creating one would add a trailing-underscore attribute that would make check_is_fitted wrongly report it as fitted; I verified an unpickled unfitted model still correctly raises NotFittedError.

Test

Added test_logger_level_preserved_after_pickle, parametrized over all three local classifiers: fit with verbose=logging.WARNING, pickle, poison the shared global logger to DEBUG, unpickle, then assert both verbose and logger_.getEffectiveLevel() are still WARNING. It fails on main (assert 10 == 30) and passes with the fix.

Closes #146

[JayeshSuryavanshi]

Note on the failing docs/readthedocs.org check: it's the docs build, not the test suite, and it appears pre-existing rather than caused by this PR. This change only adds a __setstate__ plus a test; I verified the docs example that exercises pickling (plot_model_persistence.py) runs fine on this branch, and all gallery examples pass locally except the ray-based plot_parallel_training.py (which downloads a dataset) that looks like the flaky cause. Separately I opened #170 to fix a real numpy 2.5 incompatibility (np.cross on 2-D vectors in the Venn-Abers calibrator).