Fix CatBoost and XGBoost evals_result handling

[mrlonely (mameikagou)]

Description

Fix evaluation-result handling in the CatBoost and XGBoost model wrappers.

• CatBoostModel.fit now preserves and updates a caller-provided evals_result dictionary instead of rebinding the local variable and silently discarding the output.
• Both wrappers now use None instead of a mutable dictionary as the default argument.
• Add focused, network-free regression tests for caller mutation and fresh defaults.

Root cause

CatBoostModel.fit assigned self.model.get_evals_result() to the local evals_result name. When a caller passed a dictionary, that dictionary was never populated. Both wrappers also declared evals_result=dict() in the function signature, sharing one dictionary across calls that omit the argument.

Impact

Callers can reliably collect train and validation metric histories through the same evals_result contract used by the other Qlib model wrappers. Existing calls that omit the argument remain compatible.

Validation

• Regression check against the previous implementation: 4 failures / 1 pass
• Patched focused tests: 5 passed
• Black (--target-version py38 --line-length 120): passed
• Flake8 with the repository ignore set: passed

Risk

Low. The model training calls and returned model objects are unchanged; the patch only corrects the optional output dictionary and removes mutable defaults.

[mrlonely (mameikagou)]

@microsoft-github-policy-service agree