Adapter errors shouldn't cause a reconnection

At some point, I had in a test

```cpp
request req;
req.push("BLPOP", "any", 3);

conn.async_exec(req, ignore, ...);
```

Redis responds with a NULL if the BLPOP times out, which causes `ignore` to issue an error, and that error gets propagated and cancels the connection. I don't think this is the best we can do - the connection is healthy and shouldn't be torn down here.

When parsing, I think we should distinguish between protocol violation errors (e.g. an invalid RESP3 message), which should cause re-connection; and adapter errors, which should make the request fail, but shouldn't cause re-connection. For this to work, we probably need to make adapters resilient to getting nodes after an error has occurred.