feat: async model call and match (old)

[zenodflow]

Thanks @daiwaid. The async implementation looks good to me overall. Two proposals:

• We should preserve the synchronous APIs. The sync and async should be in separate code path.
• This PR should probably wait until we figure out the batch match (multiple questions - answers in an input prompt) performance.

[zenodflow]

For naming: let's name all async methods explicitly with an async_ prefix, e.g., async_match. Its sync wrapper should name as-is.

We should get an idea of performance (tput) improvement with async first, also the overhead of asyncio loop init in the sync call.

[zenodflow]

Note that vs. #34 (query batching, async support is needed because query batching does not work well when there is tool use (e.g., browsing) or more complex match outputs.

[zenodflow]

With #34 in, we should decide the necessity for async match (async model call is still well motivated, though) -- it should further improve the speed exploiting request-level parallelism.

[daiwaid]

Continued in #37.