[bug] pipecat: prompt token count not folded, cache_write and audio details never set

[chrikrah]

Describe the bug

LLMTokenMetricsDataExtractor in openinference-instrumentation-pipecat/.../_attributes.py:544 maps pipecat's LLMTokenUsage onto the span with three gaps:

1. llm.token_count.prompt copies prompt_tokens. For AnthropicLLMService that value excludes the cache, so the span shows prompt below prompt_details.cache_read. spec/semantic_conventions.md:214-218 wants the cache counts added back.
2. cache_creation_input_tokens is never read, so prompt_details.cache_write is never set.
3. Both audio details read audio_tokens (:561, :567). LLMTokenUsage has no such field in pipecat 1.3.0 or 1.12.0. Since 1.6.0 it has input_audio_tokens and output_audio_tokens. Neither audio attribute is ever emitted.

To Reproduce

This is repro.py. The first usage is what AnthropicLLMService._report_usage_metrics builds:

from pipecat.metrics.metrics import LLMTokenUsage, LLMUsageMetricsData
from openinference.instrumentation.pipecat._attributes import _llm_usage_metrics_data_extractor as ex
anthropic = LLMTokenUsage(prompt_tokens=100, completion_tokens=40, cache_read_input_tokens=900,
                          cache_creation_input_tokens=50, total_tokens=1090)
audio = LLMTokenUsage(prompt_tokens=300, completion_tokens=120, total_tokens=420,
                      input_audio_tokens=250, output_audio_tokens=100)
for name, u in (("anthropic", anthropic), ("audio", audio)):
    a = ex.extract_from_metrics_data(LLMUsageMetricsData(processor="p", value=u))
    print(name, {k.replace("llm.token_count.", ""): v for k, v in a.items() if "token_count" in k})

$ python repro.py
# pipecat's startup INFO banner omitted
anthropic {'prompt': 100, 'completion': 40, 'total': 1090, 'prompt_details.cache_read': 900}
audio {'prompt': 300, 'completion': 120, 'total': 420}

Expected behavior

For the Anthropic usage, prompt 1050 and prompt_details.cache_write 50. For the audio usage, prompt_details.audio 250 and completion_details.audio 100.

Environment

openinference-instrumentation-pipecat at main 90c1e28, pipecat-ai 1.12.0, Python 3.12.3, Linux.

Additional context

pipecat documents on LLMTokenUsage that Anthropic and Bedrock report prompt_tokens net. OpenAI-compatible services report it gross.

Gaps 2 and 3 need no decision. I will open a pull request for them and extend test_extract_llm_usage_metrics.

@mikeldking, for gap 1, should the extractor fold by service name, or should pipecat expose a gross prompt count?