Fix observation ensemble top-k tie handling

[FlorianPfaff]

Summary

• patch observation-ensemble top-k metrics to use stable exact-k tie handling via np.argsort(-p, kind="mergesort")
• prevent tied rows from favoring higher class labels in top2_accuracy / top3_accuracy
• add regression coverage for direct helper use and metrics summaries with arbitrary prob_class_* labels

Tests

• Not run as full repo tests here: this execution environment cannot clone GitHub or install the repository dependency set.
• Locally checked the tie case in isolation: old ordering reported top-2 accuracy 2/3; patched stable ordering reports 1.0.

[github-actions]

✅MegaLinter analysis: Success

Descriptor	Linter	Files	Fixed	Errors	Warnings	Elapsed time
✅ ACTION	actionlint	22		0	0	1.17s
✅ COPYPASTE	jscpd	yes		no	no	21.76s
✅ MARKDOWN	markdownlint	101	0	0	0	2.5s
✅ PYTHON	ruff	794	0	0	0	0.61s
✅ REPOSITORY	git_diff	yes		no	no	0.08s
✅ YAML	prettier	62	0	0	0	2.44s
✅ YAML	v8r	62		0	0	10.89s
✅ YAML	yamllint	62		0	0	1.82s

Notices

📣 MegaLinter 9.5.0 is out! Discover the new features and security recommendations in the release announcement. (Skip this info by defining SECURITY_SUGGESTIONS: false)

See detailed reports in MegaLinter artifacts

Show us your support by starring ⭐ the repository

[github-actions]

Test Results

     4 files  ±0       4 suites  ±0   4m 4s ⏱️ +2s
 3 050 tests +2   3 016 ✅ +2   34 💤 ±0  0 ❌ ±0 
12 200 runs  +8  12 064 ✅ +8  136 💤 ±0  0 ❌ ±0 

Results for commit 5b25a99. ± Comparison against base commit 869a643.

[FlorianPfaff]

The implementation is already present on main; I added the missing regression coverage in e9e4bb9, so closing this conflicted PR as superseded.