The rapid evolution of 5G networks, with diverse traffic classes and demanding services, highlights the importance of Open Radio Access Networks (O-RAN) for enabling RAN intelligence and performance optimization. Machine Learning-powered xApps offer novel network control opportunities, but their resource demands necessitate efficient orchestration. To address these issues, we present OREO, an O-RAN xApp orchestrator that, using a multi-layer graph model, aims to maximize the number of RAN services concurrently deployed while minimizing their overall energy consumption. OREO's key innovation lies in the concept of sharing xApps across RAN services when they include semantically equivalent functions and meet quality requirements. Despite the NP-hard nature of the problem, numerical results show that OREO offers a lightweight and scalable solution that closely and swiftly approximates the optimum in several different scenarios. Also, OREO outperforms state-of-the-art benchmarks by enabling the co-existence of more RAN services (14.3% more on average and up to 22%), while reducing resource expenditure (by 48.7% less on average and up to 123% for computing resources). Moreover, using an experimental prototype deployed on the Colosseum network emulator and using real-world RAN services, we show that OREO leads to substantial resource savings (up to 66.7% of computing resources) while its xApp sharing policy can significantly enhance quality of service.