fix(config): read YAML config files as UTF-8

[nankingjing]

Problem

skillopt/config.py::_load_yaml opens YAML config files without specifying an
encoding:

with open(abs_path) as f:
    cfg = yaml.safe_load(f) or {}

open() therefore decodes using the platform's locale-dependent default
encoding (locale.getpreferredencoding()), not UTF-8. The shipped configs
contain UTF-8 non-ASCII characters — for example the very first line of
configs/_base_/default.yaml:

# SkillOpt default configuration — base for all environments.

The — (U+2014 em-dash), along with → and box-drawing ─ used elsewhere in
the configs, are multi-byte in UTF-8. On any host whose default encoding is not
UTF-8 (e.g. Windows with a cp936/cp1252 locale), load_config() raises
UnicodeDecodeError before training can start, breaking both entry points
(scripts/train.py and scripts/eval_only.py).

Reproduction (locale where preferred encoding is not UTF-8):

UnicodeDecodeError: 'gbk' codec can't decode byte 0x94 in position 35: illegal multibyte sequence

Fix

Open the YAML file as UTF-8. YAML is UTF-8 by spec, and the rest of the
codebase already opens text files with encoding="utf-8" (e.g.
skillopt/prompts/__init__.py, skillopt/datasets/base.py), so this just
restores parity.

-    with open(abs_path) as f:
+    with open(abs_path, encoding="utf-8") as f:

Verification

• python -m py_compile skillopt/config.py passes.
• Before the change, load_config("configs/_base_/default.yaml") raised
  UnicodeDecodeError under a cp936 locale; after the change,
  load_config(...) + flatten_config(...) succeed for the base config and
  every env config that inherits it (searchqa, spreadsheetbench, alfworld, and
  features/soft_gate.yaml).
• On a UTF-8 locale behavior is unchanged.