Support Flexible Inputs to vertexai.types.Metric

Many evaluation metrics do not require extra context provided by prompts. In fact, the prompt may confuse the metric. For example, a non_advice metric might start checking response on instruction-following confusing instructions for extensions to non_advice criteria.

My use case for evaluating responses without relying on a prompt is currently not supported.

Hardcoding an LLMMetric with only a {response} field, and

- supplying an eval_dataset with only a response column raises EvalDatasetSchema.UNKNOWN
- while supplying an eval_dataset with prompt and response columns raises INVALID_ARGUMENT

The current workarounds (supplying dummy prompt or extra 'ignore prompt' instructions to the judge) are cumbersome.

I propose adding infrastructure to support the bare minimum 'response' input, in the form of a more flexible `evaluate` -- so addressing the INVALID_ARGUMENT error.

This would be a welcome addition to the existing infrastructure supporting extra inputs.

Start by tracing vertexai.types.Metric and the evaluate path described in the issue, then inspect how EvalDatasetSchema handles response-only and prompt-plus-response columns. Reproduce the UNKNOWN and INVALID_ARGUMENT cases with the two dataset shapes. Done means a response-only metric can be evaluated without requiring a prompt, while existing extra-input behavior remains supported.

由索引模型根据 Issue 内容生成。