feat: expose cache_creation_tokens in LLM metrics

[chasleslr]

What

Surfaces prompt-cache creation (write) token counts through the metrics pipeline so downstream consumers can report on cache-creation usage.

• Adds cache_creation_tokens to LLMMetrics and wires it from the existing CompletionUsage.cache_creation_tokens at the metrics construction site.
• Adds input_cache_creation_tokens to the aggregate LLMModelUsage and accumulates it in ModelUsageCollector.collect().
• Includes cache_creation_tokens in the log_metrics LLM log line.

Why

Providers already populate CompletionUsage.cache_creation_tokens (Anthropic from cache_creation_input_tokens; AWS Bedrock as of #6663), but the value was silently dropped when LLMMetrics was constructed — LLMMetrics had no such field. So cache reads were observable (prompt_cached_tokens) while cache writes were not, on either the per-request metrics_collected signal or the aggregate session_usage_updated / session.usage signal. This surfaces already-collected-but-discarded data.

Backward compatibility

All new fields default to 0. No existing construction site or consumer changes behavior. The deprecated UsageSummary / UsageCollector is intentionally left untouched.

Testing

• tests/test_metrics_usage.py (unit): defaults-to-zero, value pass-through on LLMMetrics, and additive aggregation into LLMModelUsage via ModelUsageCollector.
• make check passes (ruff format + lint + mypy strict).

Notes

• Rebased on latest main. Provider-side emission of cache_creation_tokens on CompletionUsage is already handled upstream (Anthropic, and AWS Bedrock via fix(aws): count cached tokens in Bedrock's prompt_tokens #6663); this PR is the remaining piece that carries the value into LLMMetrics / LLMModelUsage.

Known limitations (follow-ups, intentionally out of scope)

• telemetry/otel_metrics.py emits an OTel counter for prompt_cached_tokens but not yet for cache_creation_tokens.
• For remote/cloud sessions, voice/remote_session.py (_session_usage_to_proto) cannot forward input_cache_creation_tokens until the external agent_pb protobuf gains a corresponding field.

---

🤖 Generated with Claude Code

[CLAassistant]

All committers have signed the CLA.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 2 additional findings.