Add per-input trialCount support to Eval()

[dflynn15]

Why?

Braintrust's Eval() function supports a trialCount parameter that runs each input multiple times to measure variance in non-deterministic LLM outputs. However, this setting applies globally to all inputs, which creates some (but minimal) friction in some evaluation workflows. For example:

1. Targeted Debugging is Expensive: When investigating a single flaky test case, you want to run it 10-20 times to understand the variance pattern. With global trialCount, this means running your entire suite 10-20 times, multiplying costs and wait time unnecessarily.

2. Mixed Determinism is Common: Real evaluation suites contain a mix of deterministic scenarios (math problems, factual lookups) and non-deterministic ones (creative writing, open-ended reasoning). Forcing the same trial count on both wastes resources.

3. Cost Scales Linearly: Every additional trial means another LLM API call. A global trialCount: 5 on a 100-item dataset means 500 API calls, even if only 10 items actually need variance analysis.

In order to address this, we've created a custom solution that I want to propose as a contribution. Specifically it, allows each data item to specify its own trialCount, overriding the global default. This gives users fine-grained control over where to invest their evaluation budget.

---

What?

Eval("My Project", {
  data: [
    { input: "stable query", expected: "..." },                    // Uses global (3)
    { input: "flaky query", expected: "...", trialCount: 10 },     // Override to 10
    { input: "deterministic", expected: "...", trialCount: 1 },    // Override to 1
  ],
  task: myTask,
  scores: [Factuality],
  trialCount: 3, // Global default
});

There is a corollary Python PR up to match it here: #1342