[ST-6] Harden framework instrumentation lifecycle and safety handling

[manas-fortifyroot]

Summary

This PR lands the framework-instrumentation follow-ups discovered while finishing ST-6.

It tightens LiteLLM completion-safety behavior, preserves LangChain compatibility with the installed wrapt API, improves LlamaIndex delegated-provider attribution, and fixes OpenAI uninstrument cleanup so wrappers are actually removed during re-init/test isolation.

What changed

• LiteLLM
  • make completion safety deterministic during wrapper finalization
  • add once-only completion-safety markers so late async logger callbacks become no-ops
  • preserve native OTel dedupe signaling on the FR safety wrapper span
  • improve async streaming detection coverage
• LangChain
  • switch wrap_function_wrapper(...) calls back to positional form for compatibility with the installed wrapt version
• LlamaIndex
  • stamp delegated-provider wrapper spans with gen_ai.request.model and gen_ai.system
  • tag delegated wrapper spans with fortifyroot.span.role="llm_wrapper" so backend extraction can skip them for LLMUsageEvent dedupe
  • improve Anthropic-style usage/provider attribution
• OpenAI
  • guard late streaming finalization paths from mutating already-ended spans
  • add proper dotted-method unwrapping so uninstrument() really removes Class.method wrappers in v0/v1/beta surfaces
  • add regression coverage for ended-span and uninstrument lifecycle behavior

Why

These changes came out of ST-6 framework telemetry validation and post-gate warning review:

• prevent ended-span warnings during OpenAI streaming teardown
• prevent LiteLLM async worker races from making completion safety non-deterministic
• preserve correct backend-side dedupe for LiteLLM native OTel child spans
• ensure LlamaIndex safety findings retain model/provider attribution without double counting
• keep the fork diff focused on fixes required for ST-6 validation

Verification

• LiteLLM instrumentation tests passing
• LlamaIndex safety-hook tests passing
• OpenAI regression tests for late finalization and uninstrument cleanup passing
• vendored SDK smoke/tests green after re-vendor