relay: cleanup error handling

A few changes intended to make error messages more clear, remove
duplicates, etc:

 - Don't log an error when xstream_write() fails in recover_xlog() -
   it's a responsibility of the caller. Logging it there results in
   the same error occuring twice in the log.
 - If recover_xlog() fails to apply a row and continues due to
   force_recovery flag, log the row's LSN - it might be useful for
   problem analysis.
 - Don't override relay error in relay_process_wal_event(), otherwise
   we can get 'fiber is cancelled' error in the status, which is
   meaningless.
 - Break replication if we fail to send an ack as it's pointless to
   continue then.
 - Log a relay error only once - when the relay thread is exiting.
   Don't log subsequent errors - they don't make much sense.
 - Set the relay cord name before setting WAL watcher: the WAL watcher
   sends an event as soon as it's installed, which starts xlog recovery,
   which is logged by the relay so we want the relay name to be valid.
   Note, there's a catch here: we used the original cord name as cbus
   endpoint name so now we have to pass endpoint name explicitly - this
   looks better anyway.

While we are at it, let's also add some comments to relay_subscribe_f()
and remove diag_is_empty() check as diag is always set when relay exits.

Part of #3910