…r-stork-5

- Apply black/ruff auto-fixes to _tensor.py and import ordering
- Add pylint disable=import-outside-toplevel for late imports
- Fix broken anchors #fields/#ndarrays in tensor.md (point to tensor_types.md)
- Add Backend, tensor*, tensor_annotation to test_api.py expected names

Made-with: Cursor

…r-stork-5

Made-with: Cursor

# Conflicts:
#	docs/source/user_guide/tensor.md
#	python/quadrants/_tensor.py

The qd.tensor_vec / qd.tensor_mat factories already forward needs_grad
through **kwargs (PR 3), but qd.Vector.ndarray and qd.Matrix.ndarray
themselves did not accept the keyword. With Backend.NDARRAY in play,
this meant qd.tensor_vec(..., backend=NDARRAY, needs_grad=True) raised
TypeError, breaking the documented per-backend grad contract.

Mirror the scalar qd.ndarray pattern (impl.ndarray:944-951): when
needs_grad=True, allocate a companion grad ndarray of the same dtype
and shape and wire it via _set_grad. Reject non-real element types
with a clear QuadrantsRuntimeError. Silently ignore on the python
backend, matching scalar ndarray.

Tests cover allocation + shape parity, kernel-side primal/grad
roundtrip, and the int-dtype rejection path, on both Vector and
Matrix variants and through the qd.tensor_vec / qd.tensor_mat
dispatchers.

…r-stork-5

…nsor

Follow-up to the PR-3 privatization of tensor_vec / tensor_mat. The
Vector/Matrix needs_grad tests added in this branch were calling the
now-private qd.tensor_vec / qd.tensor_mat. Update them to the public
qd.Vector.tensor / qd.Matrix.tensor classmethod surface.

No behavioural change — same factories under the hood (Vector.tensor /
Matrix.tensor delegate to _tensor_vec / _tensor_mat).

…r-stork-5

…r-stork-5

Replace the sister-test pairs in test_tensor_grad.py with
@pytest.mark.parametrize("backend", [FIELD, NDARRAY]) on a single test
function. Coverage widens for the FIELD side: we now run the
Vector/Matrix grad kernel roundtrips on FIELD as well as NDARRAY (only
NDARRAY had kernel-roundtrip coverage before; FIELD only had grad
allocation asserts).

Tests kept single-backend:
- test_tensor_vec_ndarray_needs_grad_rejects_int_dtype
- test_tensor_mat_ndarray_needs_grad_rejects_int_dtype

  These probe the QuadrantsRuntimeError raised by the new
  Vector.ndarray / Matrix.ndarray needs_grad gate. The FIELD path goes
  through legacy create_field machinery with a different error class
  and message; pinning that here would couple this branch's contract
  to upstream behaviour outside our control.

Matrix.tensor and Vector.tensor classmethods were added in this PR but
not reflected in the expected API snapshot, causing test_api[Matrix] to
fail with an extra 'tensor' attribute.

Made-with: Cursor