Fine-Grained Reward Optimization for Machine Translation using Error Severity Mappings