Fix m-LiNGAM bug for small sample size and improve documentation

[matteoceriscioli]

Good day,

While experimenting with m-LiNGAM, we observed that with small sample sizes and high percentages of missing data, logistic regression can fail when identifying the causes of missingness.

This update introduces a mechanism to increase sample availability in such cases by removing a minimal set of potential parent variables that are partially observed. Similarly, if there are not enough datapoints for the adaptive Lasso procedure (samples < features), the algorithm now attempts the procedure without bias correction (as bias correction reduces the number of available samples). Both procedures preserve the sample-limit properties as they are never executed for large enough samples, while allowing the execution on small datasets with substantial missing data.

Additionally, we:

• Fixed invalid symbols in the documentation and highlighted the code example.
• Improved the tutorial example by correcting a label and choosing a more illustrative case.
• Removed the deprecated penalty argument from the sklearn logistic regression calls to suppress warnings and avoid deprecation.

Please let me know if you have any questions or need further changes.

Best regards,
Matteo Ceriscioli

[ikeuchi-screen]

Thanks for the PR! Really appreciate it.
I’ll go ahead and merge this soon.