Add an offset argument k to diag_vec / diag_mat

[Transurgeon]

CVXPY's cp.diag(x, k) accepts a diagonal offset, and every existing CVXPY canonicalization backend supports it via the diag_vec/diag_mat LinOp data. The diff engine only implements the main diagonal, so the CVXPY-side converter has to reject k != 0:

NotImplementedError: diag_vec with k != 0 not supported in diff engine

This is currently the last engine-side atom gap blocking DIFFENGINE from becoming CVXPY's default canonicalization backend for parameter-free problems (the other, Concatenate, is CVXPY-side). An earlier attempt emulated the offset with selector matmuls in Python; that was dropped as the wrong layer — it belongs in the engine's node constructors.

Requested API

new_diag_vec(expr *child, int k) and new_diag_mat(expr *child, int k) (include/atoms/affine.h:39-40), with k threaded through the SparseDiffPy bindings py_make_diag_vec / py_make_diag_mat, which today parse a single capsule with PyArg_ParseTuple(args, "O", ...) (sparsediffpy/_bindings/atoms/diag_vec.h, diag_mat.h).

The two halves differ a lot in cost

diag_mat (extract a diagonal) is nearly free. new_diag_mat (src/atoms/affine/diag_mat.c:27) is already implemented as an index node — new_index(child, n, 1,indices, n) over indices[i] = i * (n + 1). Supporting k is just a different index list and length: for an (n, n) input, n - |k| entries at

indices[i] = (i + MAX(0, -k)) + (i + MAX(0, k)) * n;

No new sparsity machinery.

diag_vec (build a diagonal matrix) needs a vtable change. new_di/diag_vec.c:102) hardcodes an n×n output and places element i at flati * (n + 1). With an offset the output becomes m×m for m = n + |k|, and element i lands at

The Jacobian path goes through the diag_vec_alloc hook (include/utils/matrix.h:106,155, used at src/atoms/affine/diag_vec.c:54), which is implemented per matrix type:

• src/utils/sparse_matrix.c:268
• src/utils/permuted_dense.c:342
• src/utils/stacked_pd.c:285

Each needs to place child row i at the offset output row rather than