Add AlwaysRetryEOF configuration option

[twmb]

When a new connection fails with EOF, the client bubbles up a non-retryable EOF error. All other EOFs (i.e. anything after the first request) are retryable. This behavior is done because so many people run into UX issues with EOFs because they are pretty hard to diagnose (it's always configuration).

This could be undesirable behavior if you're running in an extremely loaded environment and know your configuration settings are correct. The client should provide a way to opt EOF into always retrying.