Option to configure Inference request timeout

[Legion2]

I'm looking for an option to configure request timeouts for inference requests. Either a global or a per request timeout would be nice. Currently we are experiencing many "stuck" inference requests which block the modelmesh. To resolve this we currently need to restart the model mesh when this happens. We currently don't know the root cause for this behaviour, but setting a timeout would mitigate the impact of some requests getting stuck. We have already set timeouts on all upstream components, but the upstream timeout/cancelation of the request is not propagated down to the modelmesh and so the inference requests are not canceled and consuming resources somewhere in the modelmesh. Here it would also be nice to have distributed tracing in model mesh to see where in the mesh the requests are stuck.

We are currently using v0.11.2 with triton runtime.