responses: normalize developer and instructions for Codex

[krystophny]

Summary

• normalize Responses developer content into a single leading system message
• merge instructions with developer/system content in the prompt shape Codex expects
• add targeted regression coverage for the normalization rules

Main files

• vllm_mlx/server.py
• tests/test_responses_api.py

Stack note

This PR is logically on top of PR #214 because it only makes sense once /v1/responses exists.

Until #214 lands, GitHub will show the combined diff from the stacked branch. After #214, the remaining diff should be the prompt-normalization change only.

Reviewer focus

The main behavior change is prompt shaping for strict chat templates and Codex-style requests:

• developer becomes system
• instructions and explicit developer/system content are merged into one leading system message
• duplicate or misplaced system-style content is avoided

Tests

• PYTHONPATH=/Users/ert/code/vllm-mlx /Users/ert/code/.venv/bin/python -m pytest tests/test_responses_api.py -q

[krystophny]

Closing this in favor of #214. The prompt-normalization change is now folded into the core Responses PR so reviewers can reason about the request-translation layer in one place.