fix(api): align OpenAI and Anthropic compatibility

[sufubao]

Summary

• align Chat Completions and Completions stop handling, validation, stream usage, and error envelopes
• fix Responses and Anthropic request/response bridges, including reasoning controls, strict tools, and parallel tool calls
• add focused protocol regression coverage for all four endpoints

Tests

• python -m pytest -q unit_tests/server/test_openai_stop_sequences.py unit_tests/server/test_openai_validation_error.py unit_tests/server/test_api_responses.py test/test_api/test_completion_multiple_choices.py test/test_api/test_seed_validation.py test/test_api/test_reasoning_effort.py test/test_api/test_anthropic_extra_body.py test/test_api/test_server_busy_handling.py (127 passed)
• Black, Flake8, compileall, and git diff --check
• OpenAI 2.40 and Anthropic 0.113 SDK response-model validation

The broader unit_tests/server run has 14 pre-existing failures in PD-master mocks and stale shared-memory radix-cache tests; all API-focused tests pass.