InferencePool: Signal Backend Protocol

[danehans]

Backend apps can signal the desired protocol by using appProtocol of a Service. Since inference backends, e.g. model servers, are not expected to use a Service, InferencePool should have a similar field that Gateway implementations can use to set the appropriate protocol when routing a request.

vLLM: Supports HTTP/1.1 (xref).
Triton: Supports HTTP/1.1 (REST) and gRPC inference protocols (xref).

[robscott]

This seems like a sensible addition to me. I'm not sure if this needs to be release blocking since it would be an additive field, but I'd support getting something in here.

I think the AppProtocol field in Service is suboptimal (I was also the one responsible for it, so this is on me). I'd like to take some time here to make sure we can represent this concept better in InferencePool, so would generally argue against including it in v1.0, or if we did, considering it an experimental field.

[danehans]

Moved to v1.2 milestone since it's not assigned and nor a release blocker.

[danehans]

This issue was fixed by #2162.