Make distributed execution work

[franklsf95]

Testing on a 4-node cluster with sf=100.

Changes:

• Removed the Worker class and use Ray tasks to run workers. The reason is that the current round-robin scheduling of worker executions could cause deadlock in some cases.
• Made RaySqlContext a regular class instead of an actor, since there is no need for it to be remote.

[franklsf95]

@andygrove This is good to go.