perf(openai): bound large-group scheduler candidate scans

[CPU-JIA]

Summary

• add windowed schedulable-account reads for scheduler snapshot cache and DB fallback
• bound OpenAI large-group candidate scanning with rotating page sampling instead of full-bucket materialization
• add scheduling config knobs for candidate page size / scan limit and document them
• include the Windows logger handle cleanup so full backend tests pass locally

Why

Large OpenAI groups can currently trigger repeated full-bucket ZRange + MGET + JSON unmarshal reads, followed by full in-memory candidate scans. Under 10k+ accounts this inflates CPU, heap, and GC pressure.

This change keeps the existing selection semantics, but makes the hot path operate on a bounded candidate window:

• snapshot reads can fetch only one ordered window
• DB fallback can fetch only one ordered window
• OpenAI load-aware selection rotates its starting page within a bounded ordered scan range
• each request only scores a small candidate pool instead of materializing the whole group

Validation

• go test -p 1 ./...
• go run github.com/golangci/golangci-lint/v2/cmd/golangci-lint@v2.9.0 run ./...