fix(rstar): backpropagate rewards from evaluated rollout nodes

[liuzhengyang699]

R* attaches the simulation steps to the search tree, but backpropagates only from the node where simulation started. The added descendants retain zero visits and zero value. Since trajectory_score reads the last node, completed trajectories can all score zero even when their evaluated rewards differ.

Return the evaluated node alongside the simulation reward and backpropagate from it in both the synchronous and asynchronous search. Each rollout still updates the root once, including when the simulation depth is zero.

Validation: 9 tests and 2 subtests passed with python -m pytest tests/test_rstar_backpropagation.py tests/test_conversation_logging_approaches.py -q. The regressions reproduce unscored terminal nodes in both paths before the fix; concurrent rollouts also check that the higher-reward trajectory is selected. All provider calls are mocked. Ruff passed for the new tests, and git diff --check passed.