Support num_simulations=0 in all policies.

[carlosgmartin]

With no simulations, a policy should act based on the prior. Currently, it doesn't:

• gumbel_muzero_policy fails with IndexError: index is out of bounds for axis 1 with size 0.
• muzero_policy and stochastic_muzero_policy act uniformly at random over all actions, including invalid ones.

The cause and fix for each is described below.

gumbel_muzero_policy

search traces the simulation loop body even when it runs zero times. Tracing reaches gumbel_muzero_root_action_selection, which indexes a table with one column per simulation (action_selection.py:147), so it fails with:

IndexError: index is out of bounds for axis 1 with size 0

Fix: skip the loop in search when num_simulations == 0. The existing post-search code then acts on the prior: the action is argmax(gumbel + logits) over valid actions, and action_weights is the prior restricted to valid actions.

The behavior for num_simulations >= 1 is unchanged.

muzero_policy and stochastic_muzero_policy

action_weights comes from Tree.summary().visit_probs, which falls back to 1 / num_actions when the root has no visits (tree.py:110). This ignores both the prior and invalid_actions.

Fix: when num_simulations == 0, use the softmax of the root's prior logits, which already include Dirichlet noise and exclude invalid actions.

The behavior for num_simulations >= 1 is unchanged.

Tests

Added test_gumbel_muzero_policy_without_simulations and test_muzero_policies_without_simulations. Both fail without this change and pass with it. All existing tests pass.