feat(llamaindex): emit streaming latency span attrs

[sayonfortify]

Adds FortifyRoot TTFT/STTG span attributes to the existing LlamaIndex streaming safety wrapper.

What changed:

• LlamaIndexStreamingSafety now records:
  • fortifyroot.llm.streaming.time_to_first_token_ms on the first output-bearing streamed delta
  • fortifyroot.llm.streaming.time_to_generate_ms at stream completion, only if a first token arrived
• Sync and async stream wrappers call those hooks while preserving the existing pending-item safety/flush behavior.
• Tests cover hook dispatch, integer-ms attr values, and empty-stream no-attribute behavior.

Notes:

• For LlamaIndex/OpenAI delegated calls, the backend usage row is expected to come from the child OpenAI provider span (the wrapper span is intentionally marked llm_wrapper and skipped by the extractor).
• For non-delegated LlamaIndex provider paths, these wrapper attrs provide the RDS extraction source.

Validation:

• python3 -m py_compile .../streaming_safety.py .../tests/test_streaming_safety.py
• python3 scripts/apply-fr-modification-notices.py --check
• conda py3.12 editable install of this package, plus focused timing smoke: TTFT=125ms and STTG=1125ms were set on the span.

Local full pytest caveat:

• pytest packages/opentelemetry-instrumentation-llamaindex/tests/test_streaming_safety.py -q is blocked in the shared py3.12 env before test collection because this checkout lacks the package test extras (chromadb, vector-store/cohere extras, onnxruntime). CI should run the repository test environment.