When establishing connection it was possible for `IOException` to get
propagated directly to the user without being wrapped in a
`ServiceUnavailableException`. This happened because `HandshakeHandler`
only wrapped security-related exceptions with driver exceptions. It was
problematic because rest of the codebase expects to deal with
`ServiceUnavailableException`s. For example rediscovery and retries
code can deal nicely with `ServiceUnavailableException`.
This commit fixes the problem by making `HandshakeHandler` wrap all
non-security related exceptions in `ServiceUnavailableException`.

Fatal errors terminate the transaction. It will fail to commit when
terminated and rollback will be a no-op. All outstanding errors from
previous queries should still be propagated.
Previously transaction checked it's terminated status too early and
sometimes did not propagate pending errors. This commit fixes the
problem by moving the termination check to a place where all result
cursors are fully fetched. So errors from cursors will always get
propagated.
Also added a constant for default `RoutingSettings` for convenience.

Use direct drivers to fetch cluster overview from cores, instead of
routing driver and write session. Wait for at most 60 seconds.