Add `dp --pt property train` wrapper and high-level property training API

[njzjz-bot]

Summary

Split from #5376.

Add a higher-level property-training entrypoint, covering both:

• a dedicated CLI workflow such as dp --pt property train ...
• a thin high-level Python wrapper such as PropertyTrain

The intent is to improve usability for property-model training without replacing the current DeePMD-kit training stack.

Scope

• provide a dedicated property-training CLI entrypoint on the PyTorch backend path
• add a thin Python training facade for property workflows
• continue to reuse the existing property fitting/training machinery and input conventions as much as possible
• add property-specific validation or UX improvements where they are low-risk and helpful
• provide minimal tests and a small end-to-end example

Suggested direction

• start as a thin wrapper around existing property fitting/training support
• it is acceptable for the first version to keep current JSON-style inputs rather than designing a brand-new config system
• prioritize making current capability easier to discover and invoke

Non-goals

• replacing the existing training internals
• building a separate AutoML-style workflow
• forcing broad new data adapters into the first version

Acceptance criteria

• users have a dedicated CLI entrypoint for property training on the PT backend
• users have a minimal Python-level training facade for the same workflow
• the implementation reuses existing internals rather than duplicating a new stack
• focused tests / examples cover the training entrypoint

Notes

This issue is likely higher-risk than predict/repr and may be best done after the inference-side workflow stabilizes.

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