BUG: Fix matrix_rank crash on empty matrices

[thakoreh]

Summary

Fixes #30915

numpy.linalg.matrix_rank now correctly handles empty matrices (e.g., shape (0, 3)) by returning 0 instead of raising ValueError.

Problem

Previously, calling matrix_rank on an empty matrix would crash with:

ValueError: zero-size array to reduction operation maximum which has no identity

This occurred because S.max() fails on empty arrays.

Solution

Added the same _is_empty_2d guard that other linalg functions (pinv, cond) already use. The rank of an empty matrix is mathematically 0, so we return zeros(A.shape[:-2], dtype=intp).

Testing

Added comprehensive tests for:

• Empty matrices: (0, 3), (3, 0), (0, 0)
• Stacked empty matrices: (2, 0, 3)
• Empty matrix with custom tolerance

All existing tests continue to pass.

Backward Compatibility

✓ No breaking changes - only fixes a crash case
✓ All existing functionality preserved

[seberg]

Please compare with the PR in #30422 which solved most of this, but missed a branch.

And on a general note that we should just ask always: We would prefer to know about any (AI) tool use. We haven't spelled things out, but a "sprintable" label is also to learn and e.g. fully automated PRs to solve them could also be triggered by maintainers.