Minimal example of deploying LLM with flashinfer APIs (e.g. through gpt-fast).

[sricketts]

@yzh119 proposed this in #67. I'm pulling it into its own issue to track separately, so we can close that old large tracking issue.