Fix segfault for decoupled models

[krishung5]

This PR fixed the segfault issues:

• when an exception is raised within a decoupled model. We expected users to handle error by returning an error response with a final flag like this https://github.com/triton-inference-server/server/blob/main/qa/L0_backend_python/decoupled/models/decoupled_execute_error/1/model.py#L95-L106. This fix should also cover the use case where an exception is raised by calling raise Exception("Intentional Error") without sending any final flag via response sender.
• when there are lots of concurrent requests to decoupled models, the request object may have the same address which causes issues when using that address as id to retrieve the response factory from the map. This PR removed the usage of response_factory_map. The backend will create a new response factory for each request, no matter if the request is a rescheduled one or not, and clean up the response_factory when response_sender goes out of scope in the Python model to avoid memory growth.

Added testing: triton-inference-server/server#6686

[Tabrizian]

Is it going to be cherry-picked for 23.12?

[krishung5]

Is it going to be cherry-picked for 23.12?

This is to fix a segfault so I think we could if we have not staged the container. @tanmayv25 May I have your thoughts on this?