Support fractional resource scheduling

[pang-wu]

Support fractional CPU and GPU resource scheduling. This PR actually achieve three goals:

• Support schedule multiple executors in one vCPU, with each executor has 1 spark core (for parallism, like how Spark CPU request work on Kubernetes, but no limit), using spark.ray.actor.resource.cpu config.

For more details, please refer to this RFC.

[pang-wu]

@carsonwang & team, please kindly let me know if you want a call to discuss this proposal.

[carsonwang]

Thanks @pang-wu for the work! How will the gpu config be used as Spark actually is not aware of the gpu resource?

[pang-wu]

@carsonwang To my understanding the GPU based actor scheduling/allocation will be done by Ray, spark's executor runs inside the actor. Whether the code inside Spark will actually use GPU is up to the user. But we actually want to solve the other side of the problem as well: if a cluster has GPU, Spark can still launch executor on the worker nodes for CPU only tasks using this config. Right now developers has to use mixed node cluster to run Spark job if they also want to run GPU workload in the same cluster. In most of our usecase, the Spark processing job is small, the current setup increase the setup complexity.
In terms of how the workload will actually use GPU, user can put GPU aware code inside spark functions. This should be somewhat similar to how Spark handle custom resource scheduling with GPU (correct me if I am wrong)?

[carsonwang]

LGTM