release(onnx): prepare 0.4.3 with installed artifact regressions

[LauraGPT]

Summary

Prepare the standalone funasr-onnx 0.4.3 release so already-merged wrapper fixes actually reach its users. PyPI's current 0.4.2 wheel still contains string raises and swallows ONNX Runtime import failures. The current funasr 1.4.15 wheel contains no funasr_onnx package, so upgrading funasr does not deliver these fixes.

Related issue: #3523. Keep it open for reporter verification.

Type of change

• Bug fix delivery / packaging
• Documentation
• Runtime or deployment validation

Validation

• Downloaded the actual public funasr-onnx 0.4.2 and funasr 1.4.15 wheels; verified PyPI SHA256 and archive integrity.
• RED against installed public 0.4.2: version/source checks and real constructor error paths fail (3 failures, 27 subtest errors).
• GREEN against the candidate wheel and source distribution in separate Python 3.12 environments: six installed-artifact tests each pass, including nine entrypoints for nested exporter ImportError, unrelated import failure, and failed model download; missing onnxruntime retains its original cause.
• Five source-release contracts pass. Both environments pass pip check; the clean sdist environment has no torch.
• Wheel/sdist build and twine metadata checks pass. The installed test intentionally skips in model-free pytest collection only when the optional distribution is absent; the CI script invocation still fails if the package is missing.
• CI now tests the actual installed wheel and sdist on Python 3.11 and 3.12. Local Python 3.11 and real acoustic inference are not claimed.
• Signed+DCO commit; source, distribution and verification backups retained before push.

User impact

This releases existing error-reporting fixes in the distribution users actually install. It does not change the inference algorithm or require PyTorch for ONNX inference.

Notes for reviewers

The runtime implementation changes are already on main; this PR updates the standalone version, artifact-level regression coverage, CI and separate-package installation guidance. Version 0.4.3 is a release candidate, not yet published to PyPI. No issue closure or publication claim until release and reporter verification.