SafeSlice: Enabling SLA-Compliant O-RAN Slicing via Safe Deep Reinforcement Learning