In env check, step only expected agents

[mgerstgrasser]

Why are these changes needed?

The check_multiagent_environments() precheck right now calls env.step() with an action dict containing actions for all agents in the environment. For turn-based environments that don't expects all agents to act at the same time, this could break things.

Related issue number

Closes #26413

Checks

• I've run scripts/format.sh to lint the changes in this PR.
• I've included any doc changes needed for https://docs.ray.io/en/master/.
• I've made sure the tests are passing. Note that there might be a few flaky tests, see the recent failures at https://flakey-tests.ray.io/
• Testing Strategy
  • Unit tests
  • Release tests
  • This PR is not tested :(

[mgerstgrasser]

As part of this, the behavior of env.action_space_contains() needed to change, as it was returning False if only a subset of agents were contained in the sample action dict. That case now returns True if the action for each agent submitted is contained in that agent's action space. I notice that env.observation_space_contains() already handled that case for observations, but raised a warning. We could do that for env.action_space_contains() too, although the wording of the warning message is confusing IMHO. It says "your environment returns...", which assumes that these methods are only ever called on an observation / action dict returned from the environment. Maybe it would make more sense to check for "incomplete" observation dicts and raise a warning in check_multiagent_environments() instead of the environment methods?