feat: chunk batch reflection scoring

[chiefmojo]

Problem

Reflection scoring processes all pending traces in a single batch. On stores with large episode backlogs (100+ traces) this produces a single massive LLM call that hits token limits, times out, or produces degraded output quality.

Changes

Adds a reflectionBatchSize config option (default: 20). Reflection scoring splits pending traces into chunks of reflectionBatchSize and processes each chunk sequentially, reusing the same LLM client across chunks.

• Zero/negative values fall back to unbounded (existing behavior)
• Each chunk produces its own scored output; results are merged before persisting
• No change to the scoring algorithm or output format