fix(bedrock): fold cache tokens into non-streaming converse usage

[teriyakichild]

Follow-up to #12, surfaced in downstream adversarial review of mezmo/aura#631 (finding 3): the non-streaming Converse path copied inputTokens raw and dropped cacheReadInputTokens/cacheWriteInputTokens, even though the internal TokenUsage parses them. AWS defines inputTokens as exclusive of cached tokens, so non-streaming prompt usage was understated whenever caching was active. The streaming path already folds correctly — this aligns the non-streaming conversion, with a regression test.

aura is streaming-only, so no aura behavior changes; the pin can move on its next routine bump.

[greptile-apps]

Greptile Summary

Aligns non-streaming Amazon Bedrock Converse token accounting with the streaming path.

• Includes cache-read and cache-write tokens in normalized prompt usage.
• Adds a regression test covering a cache-enabled Converse response.

Confidence Score: 5/5

The PR appears safe to merge with no actionable correctness or security issues identified.

The new non-streaming conversion matches the existing streaming arithmetic, handles absent cache counters as zero, preserves provider-reported output and total usage, and has regression coverage for cache-active responses.

Important Files Changed

Filename	Overview
rig-integrations/rig-bedrock/src/types/assistant_content.rs	Correctly folds optional cache token counts into non-streaming input usage and verifies the normalized totals with a focused regression test.

_{Reviews (1): Last reviewed commit: "fix(bedrock): fold cache tokens into non..." | Re-trigger Greptile}