Map OpenAI cached tokens in Java SDK

[Alex Jean-Baptiste (justcodebruh)]

Summary

• Map OpenAI cached-token usage into Braintrust prompt_cached_tokens metrics.
• Support both Chat Completions usage.prompt_tokens_details.cached_tokens and
  Responses API usage.input_tokens_details.cached_tokens.
• Add focused InstrumentationSemConv tests for both payload shapes.

Root Cause

The Java SDK preserved the raw OpenAI usage object, including cached_tokens,
but InstrumentationSemConv only mapped prompt, completion, total, and
reasoning token fields into Braintrust span metrics.

As a result, OpenAI responses could contain:

usage.prompt_tokens_details.cached_tokens

but the exported Braintrust span omitted:

metrics.prompt_cached_tokens

Validation

• Ran focused SDK tests:

./gradlew :braintrust-sdk:test --tests
dev.braintrust.instrumentation.InstrumentationSemConvTest --console=plain

- Built and published a local 0.3.15-DIRTY SDK artifact.
- Reran the LangChain4j OpenAiStreamingChatModel GPT-5.2 repro locally against a private Braintrust instance.

- Verified the exported Braintrust span now includes:   "prompt_cached_tokens": 9984

[Andrew Kent (realark)]

we need to add spec tests for these metrics to make sure all the openai integrations are actually mapping metrics to the semconv tool. I'll knock this out in the next day or two. Thanks for getting us started, Alex!