[pt_expt] Graph-lower .pt2 export/inference is energy-only; non-energy models lack deployment consumers

## Summary

Graph-lower `.pt2` export is restricted to energy-output models. This is not an artificial gate: for non-energy models there is currently no deployment stack that could consume a graph artifact. This issue records the missing consumers so the restriction can be lifted per output type as demand appears.

## Current state

- The graph export entry points are methods on `EnergyModel` (`forward_lower_graph_exportable` / `forward_lower_graph_exportable_with_comm` in `deepmd/pt_expt/model/ener_model.py`), and the export gate in `deepmd/pt_expt/utils/serialization.py` keys on `model_uses_graph_lower`, which requires `"energy"` in the atomic output def.
- The Python `DeepEval` graph fast path (`deepmd/pt_expt/infer/deep_eval.py`) translates energy output keys only.
- C++: `DeepPotPTExpt` is the only graph runner. There is no C++ DeepDOS in any backend (dos inference is Python-only), and `DeepTensor` (dipole/polar) has no pt_expt support even for the DENSE `.pt2` schema.
- LAMMPS consumes energy models only.

An exported graph artifact for a dos/dipole/polar model would therefore be a file nothing can load; the gate refuses to produce it.

## What lifting the restriction requires (per output type)

1. **dos / property (Python-only inference):** generalize the exportable graph lower off `EnergyModel` (output-def-driven key translation, shared with the training generalization in the companion issue) and extend the `DeepEval` graph branch to non-energy output keys. No C++ work needed.
2. **dipole / polar:** first requires dense `.pt2` support in `DeepTensor` for the pt_expt backend, then the graph schema on top. Substantially larger; only worth doing with a concrete use case.

## Suggested order

Do the training-side generalization first (companion issue), then 1, then 2 on demand.

## Context

Gate introduced on #5779; companion training-side issue: #5805.

Start with deepmd/pt_expt/model/ener_model.py, deepmd/pt_expt/utils/serialization.py, and deepmd/pt_expt/infer/deep_eval.py, then read companion issue #5805 and the existing DeepPotPTExpt runner. Define the output type to address first; done means its graph export has a deployment consumer, with non-energy keys supported in DeepEval where applicable. Dipole and polar additionally require DeepTensor dense .pt2 support.

Written by the indexing model from the issue text.