fix(server): preserve delimiters in string metadata overrides

[allenflux]

String metadata overrides such as tokenizer.chat_template=str:user: or tokenizer.chat_template=str:{{ messages[0]['role'] == 'user' }} currently fail during server model setup with ValueError: too many values to unpack. The parser splits the entire value at each = and :, including delimiters inside the string.

Split only once at the key/value and type/value boundaries so the string reaches the model constructor unchanged. Add regression tests for colons, equals signs, empty values, Unicode, and existing numeric/boolean overrides, plus a changelog entry.

Validation on macOS arm64, Python 3.12.11, with the repository's pinned native library built with GPU backends disabled:

• Original source: six string regression cases fail; two control cases pass.
• make test with PYTEST_ADDOPTS="tests/test_server_model.py tests/test_llama_chat_format.py tests/test_llama_grammar.py tests/test_llama_speculative.py": 14 passed.
• make lint: Ruff checks and formatting pass across llama_cpp and tests.

The new tests use real server settings and loader code, mocking only model construction. No model weights were downloaded; the full model/inference suite was not run. The native dependency was built unchanged.

AI assistance was used for investigation, implementation, and validation.