Add a high-level Python API for property prediction and representation extraction

[njzjz-bot]

Summary

Split from #5376.

Add a compact high-level Python API for property inference and representation extraction.

The main goal is to support notebook/script usage with a much smaller surface than directly working with lower-level internals.

Scope

• add a high-level Python facade for prediction, e.g. PropertyPredict
• add a high-level Python facade for representation extraction, e.g. StructureRepr
• keep these wrappers thin and implementation-oriented, delegating to existing internals and/or the new CLI-level workflow code where appropriate
• document the minimal public API and expected inputs/outputs
• provide focused tests and simple examples

API shape (illustrative)

from deepmd.property import PropertyPredict, StructureRepr

predictor = PropertyPredict(load_model='exp')
y_pred = predictor.predict(data=test_data)

repr_model = StructureRepr(load_model='exp')
reprs = repr_model.get_repr(data=test_data)

Exact names can still be discussed, but the deliverable should be a small, stable, script-friendly API.

Non-goals

• full training API in this issue
• broad convenience loader support for every chemistry/materials file format
• introducing a separate new model stack

Acceptance criteria

• there is a documented Python-level entrypoint for property prediction
• there is a documented Python-level entrypoint for representation extraction
• the API is thin, tested, and usable from scripts/notebooks without digging into low-level internals
• examples/tests demonstrate the intended usage

Notes

This issue should follow the CLI/inference/repr MVP work split from #5376 rather than trying to define abstractions in isolation.

Authored by OpenClaw (model: custom-chat-jinzhezeng-group/gpt-5.4)