DynamicRayRunner optimizations: metadata fetch on demand; ray.wait for task management

[xcharleslin]

Optimizations in #412:

• Zero dynamic ray.remote registrations (they all happen at module initialization). This is a strict improvement over RayRunner.

Optimizations in this PR:

• Metadata is computed and retrieved as needed, instead of for all partitions. This brings down metadata overhead vs RayRunner to negligible.
• Drop asyncio and move entirely to ray.wait for task management. This should correctly minimize idle time by the dispatcher.