Make MODEL runtimes pluggable

[ieaves]

Define a syntax that simplifies adding new runtimes.
Current runtimes:

llama.cpp
vllm
stable-diffusion
OpenVINO

[olliewalsh]

Note: there is runtime dependant logic beyond just the command line to be run (e.g health check logic, model switching apis etc...). The inference engine spec approach may not be the best method to represent this. Alternatives such as python plugins could be considered.