In my experience multi-writer is because of latency or HADR stuff - have all your data in all regions at the same time, but the method (via the tlog) seems like it sort of defeats what those sorts of systems might be able to historically do (write multiple places from the app at the same time so as to have the lowest possible chance of data loss.)
Yeah, some instances also have all the shards maintain all the state and just accept writes in your partition of values, merge replication in sql server works like this.
