We're hypothesizing that this will fix some intermittent errors we're
seeing in this test, where we get several notfound results after taking
down node 2. If two replicas end up on node 2 after we kill node 1, then
taking down node 2 would leave only one replica left, so an r=2 read
could return notfound. This matches up with our findings that subsequent
reads return the correct data (suggesting that the first read attempt is
triggering read repair).
Merging involved changing `deploy_nodes` and `build_cluster` to
support both mixed-version deployment as well as supplying
initial app.config overrides.
Updated all tests from both side of the merge to support new API
Conflicts:
src/rt.erl
src/rtdev.erl
