[testing][shortfin llm] Add soak tests

[renxida]

Our current concurrent-request tests aren't sufficient to catch most memory leaks / concurrency issues.

We probably also want some soak tests (https://en.wikipedia.org/wiki/Soak_testing), e.g. run for 30 minutes with a variable number of requests coming in throughout that period. These tests are unlikely to catch memory leaks or concurrency issues.

We should check what vLLM, SGLang, and llama.cpp do too.

(from Scott's comment on #1028)