Handle empty anomaly lists in the point metrics

[DMZ22]

Closes #527.

Problem

_point_partition derives the partition range with min()/max() over the union of the expected and observed anomalies, before the start/end overrides are applied:

edge_start = min(expected.union(observed))
if start is not None:
    edge_start = start

When neither side has any anomalies that union is empty, so every point metric raises ValueError: min() arg is an empty sequence — and because the min() runs first, supplying a range through data/start/end doesn't help either:

>>> point_confusion_matrix([], [])
ValueError: min() arg is an empty sequence
>>> point_confusion_matrix([], [], start=0, end=9)   # a range is known, still raises
ValueError: min() arg is an empty sequence

Fix

Take start/end into account first and only fall back to min()/max() when they aren't given, so a known range is enough to build the partition. If there are no anomalies and no range to fall back on, return an empty confusion matrix as you specified on the issue — there are no true or false positives and no false negatives, but the number of true negatives is unknown. That is the same (None, fp, fn, tp) shape contextual.py already returns when the true negatives can't be determined, so it keeps the 4-tuple contract the metrics unpack.

Resulting behaviour

call	before	after
point_confusion_matrix([], [])	ValueError	(None, 0, 0, 0)
point_confusion_matrix([], [], start=0, end=9)	ValueError	[10, 0, 0, 0] (all true negatives)
point_precision/recall/f1_score([], [])	ValueError	nan, via the existing zero-division handling
point_accuracy([], [])	ValueError: min() arg…	the existing "cannot obtain accuracy" error, since the true negatives are unknown
point_accuracy([], [], start=0, end=9)	ValueError	1.0

Everything with at least one anomaly on either side is untouched — point_confusion_matrix([1,2,3], [2,3,4]) still returns [0, 1, 1, 2], and [2, 1, 1, 2] with start=0, end=5.

Tests

Added test_point_confusion_matrix_empty, test_point_confusion_matrix_empty_with_range and test_point_scores_empty; all three fail on master and pass here. tests/unit/evaluation passes (36 tests), and flake8/isort are clean on the changed files.

[CLAassistant]

All committers have signed the CLA.