Add `dp property predict` MVP for high-level property inference

### Summary Split from #5376. Add an MVP for `dp property predict` as a higher-level inference entrypoint for property models. This issue should focus on making structure-based property inference easier to use on top of existing DeePMD-kit internals, especially `deepmd.infer.deep_property.DeepProperty`. ### Scope - implement a high-level `dp property predict` CLI - load a property model / checkpoint from an existing trained model - run batch inference on structure-based inputs already natural for DeePMD-kit - export prediction results in a machine-friendly format such as `npy`/`npz` - optionally support atomic outputs when the model/task naturally exposes them - provide minimal tests and a small end-to-end example ### Suggested implementation direction - wrap `deepmd.infer.deep_property.DeepProperty` - standardize model loading, batched inference, and output writing - keep input scope modest in the first version; prioritize current DeePMD-native workflows ### Non-goals - broad file-format adapter support in the first iteration - extensive CSV/reporting UX polish - notebook-oriented Python facade in this issue ### Acceptance criteria - users can run a dedicated CLI command for property prediction without manually using low-level inference internals - the command can load an existing model and produce prediction outputs for a batch of structures - outputs are written in a documented format (`npy` or `npz` is sufficient for MVP) - at least one focused test / example demonstrates the full path from structure input to property output ### Notes This issue should build on the shared CLI/property plumbing introduced under the split of #5376. Authored by OpenClaw (model: custom-chat-jinzhezeng-group/gpt-5.4)