feat(ai): Expose provider metadata as an attribute on exported OTEL spans

[dancer]

Background

Right now there doesn't appear to be a way for tracing providers to get accurate metadata if users are using Anthropic/OpenAI/Gemini's prompt caching features. The tokens used to create/read caches are returned in a providerMetadata field on the response that does not seem to be passed anywhere into the exported spans.

Summary

Adds a new ai.response.providerMetadata span attribute that exposes this response field.

Verification

Ran pnpm build and modified the anthropic-cache-control.ts example. Saw the new span attribute present:

{
  resource: {
    attributes: {
      'service.name': 'unknown_service:/usr/local/bin/node',
      'telemetry.sdk.language': 'nodejs',
      'telemetry.sdk.name': 'opentelemetry',
      'telemetry.sdk.version': '2.0.1'
    }
  },
  instrumentationScope: { name: 'ai', version: undefined, schemaUrl: undefined },
  ...
  attributes: {
    'operation.name': 'ai.generateText',
    'ai.operationId': 'ai.generateText',
    'ai.model.provider': 'anthropic.messages',
    'ai.model.id': 'claude-3-5-sonnet-20240620',
    'ai.settings.maxRetries': 2,
    'ai.prompt': `...`,
    'ai.response.providerMetadata': '{"anthropic":{"cacheCreationInputTokens":2157}}',
    'ai.usage.promptTokens': 10,
    'ai.usage.completionTokens': 456
  },
  ...
}

Tasks

• Tests have been added / updated (for bug fixes / features)
• Documentation has been added / updated (for bug fixes / features)
• A patch changeset for relevant packages has been added (for bug fixes / features - run pnpm changeset in the project root)
• Formatting issues have been fixed (run pnpm prettier-fix in the project root)

Related Issues

Fixes #7079