fix: honor task-level num_samples and warn when it has no effect

[Abelo9996]

What

Honor the task-level num_samples config field, and warn when it cannot take effect.

Why

LightevalTaskConfig exposes a num_samples field (documented as "Number of samples to generate per input"), but LightevalTask.__init__ reset self.num_samples to [1] and only extended it from sampling metrics. A task-level num_samples was therefore silently dropped, and generation stayed at a single sample unless a GENERATIVE_SAMPLING metric such as pass@k or maj@k was attached. That is the behaviour reported in #618.

Change

• Seed self.num_samples from config.num_samples (accepting a list or a bare int, keeping 1 present for base generative scoring). The request builder already takes the max, so the largest requested value wins.
• Warn when num_samples > 1 is requested but no attached metric consumes multiple samples, so a misconfiguration is loud rather than silently generating unused samples.
• Add unit tests for the default, list, bare-int, and warning cases (no model or network needed).

Open question

Per my comment on #618, this defaults the no-consuming-metric case to a warning. If you would prefer a hard error, that is a one-line change and I am happy to switch.

Scope

This covers the LightevalTaskConfig(num_samples=...) path. The metric_options={...: {"num_samples": N}} path already sets the count in Pipeline._update_num_samples; I left it untouched here and can extend the same warning there in a follow-up if useful.

Refs #618