refactor(dpmodel): use tracing-safe device helper in change_type_map

## Summary

`TypeEmbedNet.change_type_map()` still calls `array_api_compat.device()` directly when constructing and extending the identity/type-embedding matrices. Replace those calls with the tracing-safe `_array_device_or_none()` helper already used by `TypeEmbedNet.call()`.

## Motivation

Inside a JAX transformation, `array_api_compat.device()` may access `.device` on a `DynamicJaxprTracer`, which raises `AttributeError`. PR #5840 fixed the same pattern in the `TypeEmbedNet.call()` one-hot path and added coverage for the electronic-configuration path; `change_type_map()` remains the last direct use of this device query in `deepmd/dpmodel/utils/type_embed.py`.

The helper preserves the intended behavior for concrete arrays while passing `None` during tracing, allowing the active Array API namespace to choose the appropriate device.

## Scope

- Update the three device arguments in `change_type_map()` (the initial `xp.eye`, new-type `xp.asarray`, and remapped `xp.eye`) to use `_array_device_or_none()`.
- Add or extend a JAX tracing regression that exercises `change_type_map()` if this operation is expected to be supported under tracing.

Follow-up from the approved review on #5840: https://github.com/deepmodeling/deepmd-kit/pull/5840#pullrequestreview-4779484957

Related: #5736, #5840

---
Authored by OpenClaw (model: custom-chat-jinzhezeng-group/gpt-5.6-terra)

Start in deepmd/dpmodel/utils/type_embed.py at TypeEmbedNet.change_type_map(), compare its three device arguments with the _array_device_or_none() usage in TypeEmbedNet.call(). Replace the direct device queries and add or extend a JAX tracing regression for change_type_map() if tracing support is expected. Done means concrete arrays retain their device behavior and tracing no longer accesses a DynamicJaxprTracer device.

Written by the indexing model from the issue text.