fix(metrics): report effective denominator for value-dropping aggregations - #4320

Resolves #4319.

Problem

nanmean discards NaN inside aggregation, but the result entry pairs
that mean with a count over every produced value. The reported number
describes a smaller population than the adjoining count implies. The
medtext task hits this because it stores np.NAN on scorer errors.

Fix

• Aggregations that drop values are marked via excludes_nan.
• _compute_task_aggregations counts the values each excluding
  aggregation consumed and returns them alongside the metrics.
• Counts publish in sample_count on the task result (NotRequired
  on _TaskMetrics), keeping the effective denominator next to the
  number it describes. A warning names the metric and both counts.

sample_len still counts produced values. Existing warnings and the
group.py weighting TODO are unchanged.

Checks

The reporter's shape now reports the mean over the surviving values
with the matching count beside the full sample_len. New tests cover
the effective denominator and the counts emitted when nothing is
excluded, plus absence on non-excluding aggregations and the result
field.
tests/test_evaluator_utils.py, tests/test_metrics.py and
tests/test_aggregation_pipeline.py are green. Ruff is clean.

Generated by Devin