Per-model latency percentiles over the last 30 days. Only events with recorded latency are included (model calls with latency_ms set).