Add state management and queuing

[KastanDay]

State management:

• Keep track of which model(s) is in memory to help with advanced batching (NOT pure FIFO)
• Prioritization?

Queuing

• Inference queue
• Advanced batching -- when the queue contains separate requests for the same model, batch them and run all jobs requesting that model before moving onto the next model (with a max of 15-20 minutes with any one model in memory, if we have other jobs waiting in the queue. This should balance efficiency, i.e. batching, with fairness, i.e. FIFO queuing).