Allow starting clusters of Scylla servers. Chain up the next
server start to the end of the previous one, and set the next
server's seed to the previous server.
As a workaround for a race between token dissemination through
gossip and streaming, change schema version to force a gossip
round and make sure all tokens end up at the joining node in time.
Make sure scylla start is not race prone.
auth::standard_role_manager creates "cassandra" role in an async loop
auth::do_after_system_ready(), which retries role creation with an
exponential back-off. In other words, even after CQL port is up, Scylla
may still be initializing.
This race condition could lead to spurious errors during cluster
bootstrap or during a test under CI.
When the role is ready, queries begin to work, so rely on this "side
effect".
To start or stop servers, use a new class, ScyllaCluster,
which encapsulates multiple servers united into a cluster.
In it, validate that a test case cleans up after itself.
Additionally, swallow startup errors and throw them when
the test is actually used.
