fix: classify OpenAI developer messages as system context

[Gambit-Checkmate]

What & why

Closes #10.

OpenAI chat messages with the developer role currently fall through to the user segment, overstating user context and understating system instructions. Classify them with system instructions while preserving the original role, content, and turn.

Changes

• Recognize developer alongside the other OpenAI roles during format detection.
• Emit developer messages in Segment.SYSTEM.
• Add regression coverage for bare arrays and wrapped message objects, including developer-only detection and developer instructions appearing after an assistant turn.
• Document the classification in the parser docstring and changelog.

Validation

• Added the regression tests first: all 4 new cases failed before the parser change.
• python -X utf8 -m pytest: 61 passed, both with .[dev] and with .[dev,tiktoken] installed (Python 3.12 on Windows).
• ruff check .: passed.
• ruff format --check src/ctxlens/parsers/openai_chat.py tests/test_parsers.py: passed.
• git diff --check: passed.

The UTF-8 flag is needed on this Windows host because the existing HTML-output test calls read_text() without an encoding; with the host's default GBK encoding, that test raises a decoding error. No unrelated encoding behavior was changed.

AI assistance was used to implement and validate this change.

Checklist

• Focused change with a clear description
• Ran format / lint / tests (see README)
• Updated docs or CHANGELOG if relevant
• Linked any related issue (e.g. Closes #123)