[c++] Allow continued training for random forest

[Mike014]

Fixes #6143.

init_model currently fails when continuing training with boosting_type="rf".

When init_model is provided, the Python package creates an _InnerPredictor and uses its raw predictions as the new Dataset's init_score. The RF booster is then initialized before the previous model is merged into it.

At that point, RF::Init() sees a non-null init_score while num_init_iteration_ is still 0 and fails on the existing check. The previous trees would only be imported afterwards by LGBM_BoosterMerge().

This PR moves that check from RF::Init() to RF::TrainOneIter(). By then, an initial model has already been merged and num_init_iteration_ reflects the existing trees.

The existing behavior for manually-provided init_score is preserved: RF still rejects it when no initial model has been merged.

A regression test was added for continuing RF training from a saved model.

Tested with:

• test_continue_train_rf: passes (10 initial + 10 additional trees = 20)
• all continue_train tests: 5 passed
• tests/python_package_test/test_engine.py: 181 passed, 4 skipped