Beyond Hard Constraints: Budget-Conditioned Reachability For Safe Offline Reinforcement Learning