Implement Phase 1H non-stream safety parity across OpenLLMetry fork packages

[manas-fortifyroot]

Summary

This PR completes the agreed Phase 1H non-stream safety scope in the FR OpenLLMetry fork.

It adds non-stream prompt/completion safety handling for the remaining in-scope leaf invocation packages, the required framework suppression layers, and LiteLLM without requiring a LiteLLM fork.

What Changed

• Added non-stream safety coverage for these leaf invocation packages:
  • alephalpha
  • bedrock
  • groq
  • mistralai
  • ollama
  • replicate
  • sagemaker
  • together
  • transformers
  • vertexai
  • watsonx
  • writer
• Added framework-layer safety coverage for:
  • langchain
  • llamaindex
• Added a new external package:
  • opentelemetry-instrumentation-litellm
• Wired LiteLLM into traceloop-sdk instrument discovery and initialization.
• Followed the minimal-delta fork pattern:
  • substantive new safety logic lives in package-local safety.py
  • existing upstream-tracked files only have narrow helper-call wiring where possible

Why

Provider/leaf-level safety from the original 4 packages was not sufficient for the rest of the agreed non-stream scope.

In particular:

• several remaining leaf SDK instrumentations had no safety handling yet
• langchain and llamaindex can create framework spans that bypass downstream provider-level hooks
• LiteLLM had to be handled without modifying the LiteLLM repo

Fixes Included From Critical Review

This PR also includes the follow-up fixes and regressions discovered during critical self-review, including:

• Groq async non-metrics completion masking gap
• LlamaIndex inherited BaseLLM wrapper coverage gap
• Transformers keyword/nested batch masking gaps
• Watsonx multi-result masking gap
• LiteLLM direct text_completion / atext_completion
• Replicate raw-string completion leakage and broader prompt-key traversal
• Together mixed-content prompt/completion masking
• Writer mixed-content prompt masking
• AlephAlpha multi-candidate completion masking and prompt-structure preservation

Testing

Validated with:

• fr-openllmetry-py/scripts/run-tests.sh --safety

Latest clean run result:

• Packages: total=36 passed=21 failed=0 install_failed=0 skipped=15

Also revalidated the affected package slices during the review/fix passes.