llmobs-evaluations

Measures the LLMObs evaluation-metrics writer egress: append (buffer + byte sizing) and flush (makePayload plus _encode, JSON.stringify with the encodeUnicode replacer). It shares the encode path with the llmobs (span writer) bench but covers the distinct evaluations writer and its payload shape. Variants cover categorical metrics, score metrics, and reasoned metrics with non-ASCII text that exercises the encodeUnicode branch.