Add `dpa-adapt extract-descriptors` MVP for representation extraction

[njzjz-bot]

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)

[MannXo]

I'm interested in picking this up. Happy to discuss approach if needed.