Config loader uses the host locale for packaged UTF-8 YAML files

## What happens

At v0.5.0 (`4aeb436`) on Windows, Python 3.11.16 uses the default CP936 locale. Loading the packaged LeRobot config fails before any model is loaded:

```python
from vla_eval.cli.config_loader import load_config

load_config("configs/model_servers/lerobot/pi05_libero.yaml")
```

```text
UnicodeDecodeError: 'gbk' codec can't decode byte 0x80 in position 6: illegal multibyte sequence
```

Running the same command in Python UTF-8 mode succeeds, and the resolved config has `args.policy_type == "pi05"`.

## Why I think the loader is involved

The first line of `configs/model_servers/lerobot/pi05_libero.yaml` contains non-ASCII `π₀.5` text. `src/vla_eval/cli/config_loader.py` currently opens YAML with `open(path)` and no explicit encoding, so Python uses the host locale.

Is Windows host-side config validation / model `serve` intended to be supported? If so, I can send a narrow PR that reads YAML explicitly as UTF-8 and adds a CPU-only behavioral regression test.

I did not load the checkpoint or run LIBERO, Docker, or GPU code for this reproduction. I am keeping the other Windows test failures out of this issue.