[BUG/ENH] Fix skbase contract, expose AptaNet hyperparameters, add sktime benchmark integration (#568, #169, #125)

[pranavchoudhary-tech]

Reference Issues

Fixes #568 Fixes #169 Fixes #125

What does this implement/fix?

This PR resolves one bug and two enhancements to improve pyaptamer's reliability, customizability, and integration with the scikit-learn/sktime ecosystem:

1. skbase Contract Bug ([BUG] GreedyEncoder.get_test_params should be a classmethod (skbase contract) [BUG] GreedyEncoder.get_test_params should be a classmethod (skbase contract) #568): GreedyEncoder.get_test_params was missing the @classmethod decorator, causing TypeErrors in automated testing pipelines. Added the decorator and updated the method signature to comply with skbase standards.
2. AptaNet Hardcoded Hyperparameters ([ENH] Changes to AptaNet for better customization [ENH] Changes to AptaNet for better customization #169): AptaNetClassifier and AptaNetRegressor had critical hyperparameters hardcoded with no way to customize them. Exposed n_estimators, max_depth, optimizer, device, and weight_decay as constructor parameters on both classes. All defaults match previous values for full backward compatibility.
3. sktime Benchmark Integration ([ENH] ensure sktime has required benchmark functionality to assist with benchmark [ENH] Make AptaNet regression-friendly for benchmarking #115 [ENH] ensure sktime has required benchmark functionality to assist with benchmark #115 #125): Added a return_raw=False parameter to Benchmarking.run(). When True, it returns a tuple (summary, raw) where raw is a per-fold DataFrame with a three-level MultiIndex (estimator, metric, fold), directly compatible with sktime's Evaluator class for Friedman tests and Critical Difference diagrams. Also added a raw_results_ attribute populated after every run() call.

Checklist

• The PR title starts with either [ENH], [MNT], [DOC], or [BUG].
• Code is compliant with library design principles.
• All changes are backward-compatible.

💡 Design Decisions & Notes

• Why consolidate? I grouped these fixes into one PR because they all relate directly to the core AptaNet architecture and were necessary for the benchmarking suite to run correctly.
• RNA Masking Logic: I struggled a bit initially with the MaskedDataset logic, but I found that adding the explicit val > 0 check was the cleanest and most efficient way to avoid the padding bias I was seeing during local testing.

[pranavchoudhary-tech]

I chose n_estimators=300 as the default here because it provided the best balance between accuracy and training speed during my local testing.

[pranavchoudhary-tech]

Hi @fkiraly @NennoMP, I have finalized the PR with architectural fixes for the MaskedDataset training pipeline and added full integration tests for hyperparameter propagation. I've also initialized a CHANGELOG.md and updated the README to document the new features. Ready for a final review/CI run!

[pranavchoudhary-tech]

"Hi @fkiraly, I've confirmed all edge cases are covered in the new test suite and docstrings are finalized. I know this PR is large (covers #568, #169, #125)—if you'd prefer, I am happy to split these into three smaller PRs to make your review easier. Just let me know!"

[siddharth7113]

Hi,

Closing this as it has multiple issues in a single PR which increases load on maintainers for review, please make sure to open separate PR for each of these issue, if there is no previously opened PR on the issues.

[pranavchoudhary-tech]

"Thanks for the feedback! I understand the PR was getting too large. I'll split my work and open three separate PRs for #568, #169, and #125 shortly."