Add `dp property repr` MVP for representation extraction

### Summary Split from #5376. Add an MVP for `dp property repr` to expose representation extraction for property models in a stable and user-facing way. A good first milestone is to provide **descriptor-level representation export** rather than trying to settle all possible latent representation semantics up front. ### Scope - implement a high-level `dp property repr` CLI - expose descriptor / representation dumping for structure-based inputs - reuse existing descriptor-evaluation infrastructure where possible (for example, `dp eval-desc` and related internals) - export representations in a machine-friendly format such as `npy`/`npz` - provide minimal tests and a small end-to-end example ### Suggested MVP semantics For the first iteration, clearly define `repr` as a representation derived from existing descriptor-level outputs or the closest stable equivalent already available in DeePMD-kit internals. ### Non-goals - defining every possible latent/internal representation in the first version - broad downstream analysis tooling - notebook-oriented Python facade in this issue ### Acceptance criteria - users can run a dedicated CLI command for representation extraction from an existing property model - the command writes representation tensors in a documented output format - the first version has a clear and stable definition of what is being exported - focused tests / examples cover the basic extraction workflow ### Notes This issue is intentionally narrower than “all representations”. The priority is to ship a useful and well-defined MVP. Authored by OpenClaw (model: custom-chat-jinzhezeng-group/gpt-5.4)