feat: add alter column nullable to non-nullable support

[Xuanwo]

This PR will add alter column nullable to non-nullable support

---

Parts of this PR were drafted with assistance from Codex (with gpt-5.2) and fully reviewed and edited by me. I take full responsibility for all changes.

[github-actions]

Code Review

P1: Bug - Nested column validation will fail

In validate_no_nulls_before_making_non_nullable, the code uses batch.column_by_name(path) where path can be a nested path like "b.c". However, RecordBatch::column_by_name only searches top-level column names, not nested paths.

When you call scanner.project(&["b.c"]), the resulting batch schema depends on how Lance handles nested projections - the column won't be accessible via batch.column_by_name("b.c").

Suggestion: Consider using the batch's first (and only) column directly since the projection contains exactly one column:

let col = batch.column(0);

Or alternatively, extract the leaf field name from the path:

let leaf_name = path.rsplit('.').next().unwrap_or(path);
let col = batch.column_by_name(leaf_name)...

Recommendation: Add a test for nested columns

The current tests only cover top-level columns. Given that alter_columns explicitly supports nested paths (as seen in test_rename_columns at line 1324), please add a test case for making a nested nullable column non-nullable.

---

Otherwise the implementation looks correct - the validation approach of scanning all data to check for nulls before allowing the schema change is sound.

[wjones127]

+1 on the review comment. Would like to have this supported in nested columns.

[wjones127]

Add a test for nested columns for this feature.