Generating Fair Consensus Statements with Social Choice on Token-Level MDPs