Fix infeasible torch validation splits

[FlorianPfaff]

Summary

• Extend the existing torch fit guard to cover additional torch-backed adaptation decoders when a requested stratified validation split is too small.
• Restore the configured validation_fraction after the fallback fit path.
• Add regression coverage for a 16-class, two-rows-per-class case with validation_fraction=0.1.

Tests

• Not run locally: this environment cannot clone GitHub or install the repository dependency set. Added focused pytest regression coverage in tests/test_torch_weight_validation.py.

[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	21.28s
✅ MARKDOWN	markdownlint	101	0	0	0	2.52s
✅ PYTHON	ruff	796	0	0	0	0.61s
✅ REPOSITORY	git_diff	yes		no	no	0.13s
✅ YAML	prettier	62	0	0	0	2.31s
✅ YAML	v8r	62		0	0	9.95s
✅ YAML	yamllint	62		0	0	1.62s

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 18s ⏱️ +27s
 3 054 tests + 3   3 018 ✅ +1   36 💤 +2  0 ❌ ±0 
12 216 runs  +12  12 072 ✅ +4  144 💤 +8  0 ❌ ±0 

Results for commit 4833597. ± Comparison against base commit f2f8d87.

[FlorianPfaff]

Applied the relevant fix and regression coverage directly to main in follow-up commits (5d8b442, 3a33044), so closing this conflicted PR as superseded.