[shortfin_apps.llm] Batcher-level generation tests

[renxida]

We need integration tests that test to make sure batching does not change generation results. Previously, we've had issues like the one fixed by #873 where improper masking caused

• token corruption due to a batch having more than 1 request
• token corruption due to a batch being not full

Our current overall shortfin llm server integration tests weren't able to reliably expose them in a replicate-able way because our concurrency tests go through the http endpoint, where the timing in which the requests come in causes differences in how they are batched.