fix: preserve OpenAI max_new_tokens

[Genmin]

Summary

• remove the silent 4096 cap from OpenAI-compatible simple and chat payload builders
• use the requested max_new_tokens for reasoning-model max_completion_tokens instead of hardcoding 5000
• add payload-level regression coverage for simple, chat, and reasoning-model paths

Fixes #1317

Testing

• uv run --with pytest python -m pytest test/models/test_openai.py -q
• uv run --with ruff ruff check lmms_eval/models/chat/openai.py lmms_eval/models/simple/openai.py test/models/test_openai.py
• git diff --check