Avoid panic on connection failure

[jaymell]

This code is not unreachable, so let's handle the error instead of panic.

[nihohit]

How did you reach that code?

[jaymell]

I actually saw it while doing a bit more local testing of #968. If I manually terminate cluster instances to test recovery, I will occasionally see this. It's also fairly easy to force by doing something egregious like adding a sleep in the auth function. Anyway, clearly it's not unreachable code 😆.

[NathanFlurry]

If the MultiplexedConnection panics, then all future commands also panic. This means that one bad AUTH and connection (and therefore the entire server) is unusable until it's restarted.

Our Redis hosting provider triggers this panic any time they enter a maintenance window. I recommend cutting a release with this fix soon. :)

---

I can consistently reproduce this issue (before this is merged) by:

1. Adding a timeout in the authenticate function to give us time to inject a fault

2. Add a timeout after the authentication so the driver future finishes first

3. Open a connection and kill the Redis server during the authenticate function.

I tested this same code with this PR, can confirm that the connection recovers correctly now.