Semaphore for model deletion and model processing.

[mattjhawken]

Currently models can be taken offline, either due to an expired job or disconnected worker, and any requests still being processed by the model will throw key errors. (eg. _handle_forward in torch_node.py).