Auto selection of inference engine

[AlexCheema]

Right now, inference engines are not interoperable e.g. you can't run one machine on MLX and another on tinygrad - they need to both be using the same inference engine.

We should automatically handle the selection of the inference engine. We can still provide a manual override with --inference-engine but the default should be automatically detecting the best inference engine that is compatible with all devices in the network.