[RLlib] Faster remote worker space inference (don't infer if not required).

[sven1977]

If an env is not constructed on the local worker (which is the default behavior), observation- and action spaces must be inferred from one of the remote workers. This causes the Trainer construction to take a considerable longer time (a few seconds) than w/o this step (ray.get).
This PR avoids this remote worker space lookup in case:

• we do have an env on the local worker (create_env_on_driver=True).
• observation AND action spaces are given manually via the config's "observation_space" and "action_space" keys.

Why are these changes needed?

Related issue number

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 :(

[sven1977]

@RuofanKong ^^

[avnishn]

LGTM

Is there a small test that you want to add here first potentially? @sven1977

[sven1977]

Good point! I'll add this to test_trainer.py.

[sven1977]

Test case added.