Model containers should have a choice of running in batch or single-input mode

[dcrankshaw]

Currently, we expect all deployed models to be a function from a list of inputs to a list of outputs. This allows models the ability to do batching, but it makes the function signature of models more complicated. For simple models that want to operate on a single-input at a time, users have to remember to wrap the input in for-loop.

Instead, we should support deploying two types of models: batch and single-input. Batch models are still given the full available batch at once. Single-input models would be a function from a single input to a single output, and the RPC server embedded in the model container would call the model in a for-loop transparently to the user.

This should make the model deployment process simpler in many common cases.

[dcrankshaw]

Based on discussion in #395, we decided that rather than supporting single-input mode, we should improve our documentation and debugging capabilities to help users use the batching functionality in Clipper correctly. #394 is a good first step in that direction.