Add backend routing, transform caching, and sklearn estimator caching

[robosimon]

Summary

This PR makes iTuna significantly smoother for large iterative workflows:

Backend routing

You can now route different operations to different backends, matching how users actually run experiments:

# Expensive estimator fits go distributed
config.register_backend_route(method="fit", model_class=ConsistencyEnsemble, backend="disk_cache_distributed")
# Cheap consistency transform fits stay local
config.register_backend_route(method="fit", model_class=ConsistencyTransform, backend="disk_cache")

Transform caching

ConsistencyTransform.fit is now cached in disk-cache workflows, so consistency alignment is never recomputed for identical inputs. The distributed backend also uses a local fast-path when all requested models are already cached, skipping redundant sweep registration.

sklearn estimator caching (ituna.sklearn)

New module for caching standalone sklearn estimators outside of ConsistencyEnsemble:

from ituna.sklearn import cached, enable_global_cache

# Cache a single estimator instance in-place
cached(my_estimator, cache_dir="./cache")
my_estimator.fit(X)  # cached on first call, loaded from disk on subsequent calls

# Or enable globally for all estimators of a class
enable_global_cache(FastICA, cache_dir="./cache")

Patched estimators pass sklearn compliance checks (check_estimator).

Changes

• ituna/config.py: register_backend_route, remove_backend_route, clear_backend_routes, resolve_backend_route
• ituna/_backends/: lazy backend resolution per operation; disk-cache transform fit caching; distributed fast-path for fully cached reruns
• ituna/sklearn.py: new module — cached(), enable_global_cache(), disable_global_cache(), get_global_cache_status()
• ituna/_cache_guard.py: recursion guard preventing double-caching when patched estimators are used inside ensembles
• Docs: updated backends tutorial with routing examples, new sklearn caching guide, CEBRA best-practices demo
• Tests: test_backend_routing.py, test_sklearn_cache_wrapper.py, updated disk-cache and distributed tests