Fix memory leak from lru_cache on _FanovaTree instance methods - #6859

Motivation

_FanovaTree._is_node_leaf, _get_node_value, and five other methods in optuna/importance/_fanova/_tree.py are decorated with @lru_cache(maxsize=None) directly on the class. Decorating an instance method this way keys the cache on self as well as the other arguments, so every _FanovaTree instance ever constructed stays reachable through the cache for the life of the process, even after the caller has dropped every other reference to it.

_FanovaTree is constructed once per estimator in the random forest fitted inside _Fanova ([_FanovaTree(e.tree_, search_spaces) for e in self._forest.estimators_]), so every call to get_param_importances() builds a batch of these and normally expects them to be freed once the importances are computed. With maxsize=None, they instead leak for good, along with whatever self._tree/self._statistics/self._search_spaces they're holding onto — this adds up in any long-running process that computes feature importances repeatedly (a dashboard, a service, a notebook running many studies).

Description of the changes

Moved the caching from class-level decorators to per-instance caching, by wrapping the bound methods in __init__ instead of decorating them at class scope:

self._is_node_leaf = lru_cache(maxsize=None)(self._is_node_leaf)

This keeps the exact same caching behavior for the lifetime of a _FanovaTree instance, but the cache (and everything it holds) is now freed along with the instance instead of outliving it.

Added a regression test (test_tree_is_garbage_collected) that constructs a tree, exercises the cached methods, drops the only reference, and asserts the instance is actually collected. It fails against the old code and passes with this fix. All existing tests in test_tree.py pass unchanged (40 passed), and ruff check/ruff format --check are clean on both changed files.