Fix three critical catboost engine bugs

[viv-analytics]

Summary

• objective alias crash: process_loss_function() was checking only for "loss_function" before auto-injecting a loss. If the user passed objective via set_engine(), bonsai still injected loss_function, triggering CatBoost's hard C++ error: "Only one of the parameters [loss_function, objective] should be initialized." Fixed by extending the guard to c("loss_function", "objective").

• Synonym alias collision: CatBoost hard-errors when synonym aliases (n_estimators, eta, max_depth, etc.) are passed alongside the canonical parameter name that bonsai injects by default. Added check_catboost_aliases() + catboost_aliases table (modelled on the existing check_lightgbm_aliases()) that errors early with a helpful message directing users to the parsnip argument instead.

• stop_iter non-functional: train_catboost() never passed a test_pool to catboost.train(), so the overfitting detector (od_type = "Iter") had no evaluation data and silently did nothing. Added a validation argument (mirrors LightGBM's pattern): when early_stopping_rounds is set, a held-out pool is built and passed as test_pool. When validation is omitted the training data is reused as the evaluation set (same safe fallback as LightGBM).

Test plan

• catboost does not inject loss_function when objective is already set — regression and multiclass with objective = pass without error
• catboost errors on synonym aliases passed to set_engine() — n_estimators and eta both error with "alias for a main model argument"
• stop_iter is functional when validation is supplied — direct API with validation = 0.2, fallback with no validation, and parsnip interface all pass
• All pre-existing catboost tests continue to pass (60 pass, 2 pre-existing snapshot skips unchanged)

🤖 Generated with Claude Code