fix: Windows MinerU parsing — subprocess UTF-8 decoding & render-thread crash

[1939566682]

## Summary
Two Windows-specific fixes for the local MinerU parsing engine.

- `local.py`: set `encoding="utf-8", errors="replace"` on the MinerU
  `subprocess.Popen` call. Previously `text=True` decoded MinerU's UTF-8
  progress output with the system locale; on Chinese Windows that is GBK,
  causing `UnicodeDecodeError: 'gbk' codec can't decode byte 0x8b` and
  aborting every parse.
- `models.py`: inject `MINERU_PDF_RENDER_THREADS=1` in `model_env_overrides`.
  MinerU 3.x crashes with `ucrtbase.dll 0xc0000409` heap corruption on
  Windows when >1 PDF render thread runs (MinerU issue #5033). Forcing 1
  avoids it with negligible impact on single-doc parsing.

## Test plan
- [ ] Parsed a multi-page Chinese PDF on Windows 11 (zh-CN), Python 3.13,
      MinerU 3.4.4 — no decode error, parse completes.
- [ ] Long-running parse (>30 min) no longer crashes with 0xc0000409.

## Notes
Verified on Windows 11 zh-CN / Python 3.13 / MinerU 3.4.4.