Add safety attribution metadata to provider hooks

[sayonfortify]

Summary

• Adds a shared build_safety_metadata helper for raw LLM context attributes.
• Passes gen_ai.system, gen_ai.request.model, and gen_ai.response.model through safety hooks for OpenAI, Anthropic, Bedrock, Google GenAI, and LiteLLM where the wrapper already has that context.
• Keeps routing/billing provider canonicalization out of the SDK/fork; backend remains the source of provider-role enrichment.

Tests

Check	Result
python -m py_compile on modified fork files	Passed
pytest packages/opentelemetry-instrumentation-fortifyroot/tests/test_safety.py -q	7 passed
pytest packages/opentelemetry-instrumentation-openai/tests/traces/test_safety_hooks.py -q	5 passed
pytest packages/opentelemetry-instrumentation-anthropic/tests/test_safety_unit.py -q	16 passed
pytest packages/opentelemetry-instrumentation-bedrock/tests/test_safety_hooks.py -q	18 passed
pytest packages/opentelemetry-instrumentation-litellm/tests/test_safety_hooks.py -q	27 passed
pytest packages/opentelemetry-instrumentation-google-generativeai/tests/test_safety_hooks.py -q	11 passed
git diff --check	Passed

Tests were run with PYTHONPATH pointing at local package dirs in the py3.12 conda env.