Adds FortifyRoot streaming-latency span attributes to the LiteLLM streaming safety wrapper.

What changed:

• sync and async LiteLLM streaming wrappers set:
  • fortifyroot.llm.streaming.time_to_first_token_ms
  • fortifyroot.llm.streaming.time_to_generate_ms
• first token is detected only from output-bearing chunks (delta.content, message.content, text/content); role-only or empty chunks do not start timing.
• empty streams leave both attrs unset so downstream RDS fields remain NULL, not fake 0 ms.
• preserves existing streaming safety masking, accumulation, status, context detach, and span-end lifecycle.

Tests:

• conda py3.12: python -m pytest packages/opentelemetry-instrumentation-litellm/tests/test_safety_hooks.py -q → 31 passed
• conda py3.12: python -m ruff check touched files → pass
• python3 -m py_compile touched files → pass
• git diff --check → clean

Vendoring note:
Per internal-sop/openllmetry/VENDORING_SOP.md, ocelle-py vendoring is intentionally not done in this PR. After this merges, cut the next fr-v0.52.6.N tag on the fork and run ocelle-py/scripts/vendor.sh against that tag.