docs(pt): sync DPA4 torch.compile version text with supported allowlist

## Problem

The DPA4/SeZM `torch.compile` runtime support was expanded, but several user-facing/source documentation strings still say that compilation requires `torch==2.11`.

The current runtime gate in `deepmd/pt/utils/compile_compat.py` is:

```python
SUPPORTED_COMPILE_TORCH = ((2, 11), (2, 12), (2, 13))
```

so PyTorch 2.11.x, 2.12.x, and 2.13.x are explicitly supported. PyTorch 2.13 support was added by #5910.

However, a repository search for `torch==2.11` still finds stale compile-version wording in:

- `deepmd/utils/argcheck.py`: `doc_use_compile` says `Requires torch==2.11.`
- `deepmd/pt/model/model/sezm_model.py`: the SeZM model docstring says the compile path `requires ``torch==2.11```
- `doc/model/dpa4.md`: the `DP_COMPILE_INFER` entry says it has the same `torch==2.11` requirement as `model.use_compile`

`pyproject.toml` also contains a `torch==2.11.0` CPU CI/build pin, but that appears to be an intentional environment pin rather than a statement of the compile compatibility contract and should not necessarily be changed as part of this issue.

## Expected

Keep the user-facing documentation and source docstrings in sync with `SUPPORTED_COMPILE_TORCH`, ideally avoiding another hard-coded version list where practical so future compatibility additions do not leave stale docs behind.

Related: #5910, #5935, #5766

---
Agent: ChatGPT
Model: GPT-5.6 Sol