Fixes around creating new connections

[piotrp]

A few fixes separated into a few commits:

• pass error from socket when connector fails to select a database, so that "timeout" error isn't hidden away from the user (without this we can get a connection with closed socket and not know about that until we try to use it, it may happen when connection gets broken outside of nginx)
• slave connection: use set_keepalive only when we are sure that Sentinel responded correctly - it may have timed out, and we don't want to put such connection in pool
• master connection: use set_keepalive always when Sentinel responded correctly, regardless of result of connecting to indicated Redis instance

[pintsized]

Hey, thanks for this. Seems these changes break the tests though, could you try running them on your side? You'll need to make start_redis_instances, followed by make test.

Here's what I'm getting: https://gist.github.com/pintsized/bff6c6beaa4799762788a32fcd1a8884

[piotrp]

Oops... sorry for that, fixed.

There were two issues, both in my first commit:

1. I removed logging of an error message when AUTH command failed, and there was a test that looked for it.
2. My original fix to error returned from :connect didn't account for Sentinels: :connect tries to set database, which is an unsupported operation in Redis Sentinel, and possibly that's why error handling of that call was originally omitted.

[piotrp]

Hello, @pintsized, were you able to look into this MR?

[pintsized]

Hey, I'm really sorry I seem to have totally forgotten about this. I managed to get travis CI working on this repo, if you rebase it should pick that up and build your pr.

Otherwise, yep LGTM. Looks like there's a 2-space indentation on line 363 if you feel like making it perfect ;)

[piotrp]

Invalid formatting is a perfectly valid issue, rebased and fixed :)