Avoid requesting policy actions after terminal timesteps

[sylvesterkaczmarek]

Summary

Avoid requesting another population action after the substrate has already returned a terminal timestep.

evaluation.run_episode currently sends the terminal timestep to the population and then unconditionally calls await_action(), even though that action can never be consumed because the episode is over.

This change:

• still forwards the terminal timestep to the population
• skips await_action() once the terminal timestep has been observed
• leaves action collection unchanged for non-terminal timesteps
• adds focused regression coverage

Testing

Added a regression test for a one-step episode and verifies that await_action() is called only for the initial non-terminal timestep, while both the initial and terminal timesteps are still delivered to the population.

[jzleibo]

Can you elaborate on why this is needed? The change makes sense of course, but I've often taken a "if it's not broken, why fix it?" sort of attitude to this kind of thing. Is this fixing a specific issue on your end? I'm happy to approve if so. But if it's just because it "makes more sense this way" then maybe better to leave it alone? What do you think?

[sylvesterkaczmarek]

Can you elaborate on why this is needed? The change makes sense of course, but I've often taken a "if it's not broken, why fix it?" sort of attitude to this kind of thing. Is this fixing a specific issue on your end? I'm happy to approve if so. But if it's just because it "makes more sense this way" then maybe better to leave it alone? What do you think?

Good point. This came from code inspection rather than a concrete user-facing or production failure. Population.reset() clears or cancels pending action futures before the next episode, so I don't have evidence that the extra terminal action causes an actual correctness issue. I agree it is better not to change the behavior without that evidence, so I'll close this PR. Thanks for the review.