Define rank-deficiency behaviour for zero-column designs

[shawcharles]

Problem

_detect_rank_deficiency() currently assumes the QR diagonal has at least one entry. For a design matrix with shape (n, 0), pivoted QR returns an empty diagonal and the helper indexes r_diag[0], raising an IndexError.

That leaves the helper without an explicit contract for a valid but empty design matrix.

Proposed behaviour

Treat an (n, 0) design as:

• rank 0;
• no dropped columns;
• empty pivot array.

This matches the mathematical rank of an empty-column matrix and avoids forcing callers to special-case a design that has no columns to drop.

Proposed scope

• Add a direct regression test for _detect_rank_deficiency(np.empty((n, 0))).
• Add a small early return for k == 0 before pivoted QR.
• Avoid broader rank-deficiency, OLS, or estimator behaviour changes.

Validation

• PYTHONPATH=. DIFF_DIFF_BACKEND=python pytest tests/test_linalg.py::TestNumericalStability::test_rank_detection_zero_column_matrix_returns_empty_contract -q
• PYTHONPATH=. DIFF_DIFF_BACKEND=python pytest tests/test_linalg.py::TestNumericalStability -q
• PYTHONPATH=. DIFF_DIFF_BACKEND=python pytest tests/test_linalg.py -q
• PYTHONPATH=. DIFF_DIFF_BACKEND=python .venv-py39/bin/python -m pytest tests/test_linalg.py::TestNumericalStability::test_rank_detection_zero_column_matrix_returns_empty_contract -q
• git diff --check