Fix: rezero buffers variable action space

[SteamMachinist]

No description provided.

[puyuan1996]

Thanks for the PR. I agree with the motivation to handle variable action spaces correctly, but I think this change may break the existing child_visit_segment contract.

For varied_action_space, the current non-reanalyzed target path treats child_visit[current_index] as a compact distribution aligned with legal_actions, and expands it back to the full action space later. This PR writes the expanded full-action policy back into child_visit_segment. As a result, when the same segment is later used by the non-reanalyzed path, it will be expanded again incorrectly.

For example, with action_space_size=5, legal_actions=[1, 3], and compact distribution [0.25, 0.75], this PR stores [0, 0.25, 0, 0.75, 0]. Later _compute_target_policy_non_reanalyzed() will read distributions[0] and distributions[1] for legal actions 1 and 3, producing an incorrect target.

I think we should either:

1. keep child_visit_segment compact and only return full-size target_policies, or
2. change all consumers of child_visit_segment to treat it as full-size, update the docs/comments, and add regression tests.

Could you add a unit test covering ReZero + varied_action_space, especially the path where a reanalyzed segment is later sampled through the non-reanalyzed target computation?