[amkkoussay]

Fixes #804

What this does

The markdown results summary (make_results_table) now shows a Count column with the number of samples that were actually evaluated for each task, exported by the evaluation tracker as num_samples alongside results.

Design choices

• Source of the count: the samples actually logged per task (DetailsLogger.details) — i.e. the filtered/evaluated set behind the metrics, not the raw dataset size. This follows the concern raised on the issue thread about counts being misleading when filtering is applied.
• Placement: the count appears once per task, on its first metric row; the remaining metric rows stay blank so the table stays readable.
• Backward compatibility: results dicts without counts (e.g. loaded from older runs) render a blank cell — no key is required and nothing breaks.

Changes

• src/lighteval/logging/evaluation_tracker.py — export num_samples in both results-dict builders (results property and generate_final_dict).
• src/lighteval/utils/utils.py — render the Count column in make_results_table.
• tests/unit/utils/test_utils.py — new unit tests for the column.

Tests

• Reproduction of the issue's example: before the change the table had no count; after the change the Count column shows the evaluated samples.
• tests/unit/utils/test_utils.py: 9/9 pass (6 existing + 3 new covering: count on the task's first metric row only, missing counts, partial counts).

Happy to adjust the key name, column position, or the source of the count if you prefer a different design.