Improve robustness of numeric data distribution visualization

## Summary

The numeric data distribution helper (`plot_numeric_distributions`) introduced in PR #145 works well for basic use cases, but there are several opportunities to make it more robust and user-friendly for ML analysis workflows.

## Tasks

- Add explicit validation for the `columns` argument:
- Ensure all requested columns exist in `df.columns`.
- Optionally enforce or warn when non-numeric columns are requested.
- Raise clear `ValueError` messages for invalid columns.
- Improve subplot layout behavior:
- Avoid creating unused extra axes for small numbers of columns (e.g. 1 column resulting in a 1x2 grid).
- Keep the layout readable for larger numbers of numeric features.
- Document headless usage considerations:
- Clarify in the docstring how to use the function safely in CI or non-GUI environments (e.g. using a non-interactive matplotlib backend).
- Consider re-exporting `plot_numeric_distributions` from the `apps/ml/visualizations/__init__.py` module if this is intended as the public API surface.

## Acceptance Criteria

- Calling `plot_numeric_distributions` with invalid or non-existent column names produces clear, actionable error messages.
- Subplot layouts are sensible for 1, 2, and many numeric columns.
- The docstring documents headless/CI usage expectations.
- Existing tests continue to pass, and additional tests are added where helpful (e.g. column validation).

Start at plot_numeric_distributions, then inspect apps/ml/visualizations/__init__.py and the existing tests around the helper. Check invalid and non-numeric column cases, subplot layouts for 1, 2, and many columns, and the docstring's headless-use guidance; run the existing test suite and add focused tests where helpful. Done means clear ValueError messages, sensible layouts, documented CI usage, and passing tests.

Escrita pelo modelo de indexação a partir do texto da issue.