CCM-33842: Normalize OpenAI and Anthropic GenAI usage attrs

[shreyas70]

Summary

• Normalize OpenAI and Anthropic client spans to assert/use gen_ai.provider.name instead of the older gen_ai.system attribute.
• Populate dotted cache token attributes when provider usage details are available:
  • gen_ai.usage.cache_read.input_tokens
  • gen_ai.usage.cache_creation.input_tokens
• Populate gen_ai.usage.reasoning.output_tokens when reasoning token details are available.
• Add test assertions for OpenAI and Anthropic semantic-convention fields.

Test plan

• .venv/bin/python -m pytest test/instrumentation/openai/openai_instrumentation_test.py -> 19 passed
• .venv/bin/python -m py_compile src/harness_sdk/instrumentation/anthropic/__init__.py src/harness_sdk/instrumentation/openai/__init__.py

Notes

• Anthropic focused tests could not be run in this local venv because the PyPI opentelemetry-instrumentation-anthropic package installed here does not include opentelemetry.instrumentation.anthropic.messages_extractors, which this repo's existing Anthropic wrapper imports.
• Streaming cache/reasoning accumulation is not changed in this PR.

Made with Cursor