Pin UTF-8 encoding when reading the shadow model's config.json

[dltsum]

Problem

ShadowModel._load_shadow_backbone reads the shadow checkpoint's config.json with the locale's preferred encoding:

config_file = cached_file(config.shadow_model, "config.json")
with open(config_file) as f:
    raw_config = json.load(f)

On Windows with a non-UTF-8 locale (e.g. cp936/GBK), a config containing non-ASCII content crashes:

UnicodeDecodeError: 'gbk' codec can't decode byte 0xa4 in position 52: illegal multibyte sequence

A realistic source: a locally fine-tuned checkpoint whose config.json carries a CJK/emoji path in _name_or_path and was saved with ensure_ascii=False.

Reproducer (Windows 11, cp936, Python 3.14)

Write {'model_type': 'bert', '_name_or_path': 'D:/模型🤗hub'} as UTF-8 JSON into a local checkpoint dir and call _load_shadow_backbone → UnicodeDecodeError before this change; the read succeeds after it (the call then proceeds to model loading as expected).

Fix

Pin encoding="utf-8" on the read.

🤖 Generated with Claude Code

[BenjaminBossan]

Please fold this PR into #3771. Similarly, if you find this kind of error elsewhere in the code, please fold it into one PR.

[dltsum]

Folded into #3771 as commit 216bc6a (same change, plus the requested regression test).