Support streaming real-time prediction

[andreasjansson]

Currently Cog requires you to upload an input file, then it's processed and results are returned. But there are cases when you might want to stream a continuous input to the model. For example, if you have a model that does audio event detection, you might want to display the current event as it happens.

[sinopec]

the same requirements,but seems not easy, This means you need to have exclusive access to the entire model, which seems to be inconsistent with the replicate business model.