[shortfin_apps.llm] Unit tests and smaller-chunk integration tests

[renxida]

We need more unit tests, and, more importantly, component-integration tests at a level between unit and whole-server integration.

I have been focusing almost exclusively on our integration tests, but this leaves me often wondering what went wrong when the one big integration test fails.

We have been over-relying on the overall integration tests in shortfin_apps. We need defense-in-depth with unit and component tests for, among other things:

• the batcher
• generate.py where we manage the lifecycle of a generation request, maintain cache allocations, and store the tokens generated so far for each request
• config management

A lot of the recent failures in the integration test took WAY too long to debug because we're missing these.

E.g. the trie-cache related incomplete-batch issue that would have been caught if we had component tests that directly sent requests to the batcher (rather than through http) and tested batches of various levels of full-ness.