Runtime env_vars cannot be passed down to executors

[Deegue]

When we use ray.init(runtime_env={"env_vars":"xxx"}), executors cannot get "env_vars" while it is critical for executors in many scenarios.

It is because RayAppMaster will raise a new job which is separated from the worker pod of RayDPSparkMaster.

[g-clef]

I'm running in to this also. Would like to be able to run experiments with different dependent libraries (and send them to the workers with an env_var) without having to change the base image of my ray cluster.