feat: support per-test variable substitution in eval input

[christso]

Summary

Add native eval-file support for per-test variables and template substitution in input / related test fields, similar to promptfoo's vars + {{name}} pattern.

Motivation

agentv import promptfoo currently has to eagerly render prompt templates into concrete strings because AgentV eval YAML appears to support only:

• environment interpolation via ${{ VAR }}
• grader prompt template variables like {{input}}, {{output}}, {{criteria}}

It does not appear to support promptfoo-style per-test variable substitution for eval inputs such as:

tests:
  - vars:
      question: What is the capital of France?
    input: "Answer clearly: {{question}}"

That makes imported suites less idiomatic and loses the original template structure.

Current behavior

AgentV eval files accept concrete input strings or message arrays. Environment interpolation works with ${{ VAR }}, but that is config/env-based, not per-test data-driven substitution.

Proposed direction

Support a minimal, explicit per-test templating model for eval files, for example:

tests:
  - vars:
      question: What is the capital of France?
    input: "Answer clearly: {{question}}"

Scope could include:

• tests[].vars as a JSON/YAML object
• substitution in input string shorthand and message content strings
• maybe criteria / expected_output strings if that fits the model
• preserve current ${{ ENV_VAR }} behavior separately

Why this matters

• makes imported promptfoo suites cleaner
• reduces duplication in eval YAML
• enables more declarative eval authoring for humans and AI agents

Notes

This is distinct from:

• environment interpolation (${{ VAR }})
• llm-grader / composite prompt template variables ({{input}}, {{output}}, etc.)

Those already exist, but they are not a substitute for per-test input templating.