[fdmath] correct Gridlock Yee spacing order

Use center-to-center intervals for forward E derivatives and base-cell widths for backward H derivatives in the Gridlock examples. Document the distinction from meshes defined directly by forward intervals and fix the random fixture's backward averaging direction.

Add coordinate-derivative regression coverage for graded periodic and nonperiodic Gridlock meshes using both functional and sparse operators.

Validation: 1689 tests passed, 586 skipped; no new Ruff findings in changed Python files.

docs/api/fdmath.md

@@ -1,5 +1,18 @@
 # fdmath
 
+For Gridlock's default E-field Yee shifts, construct solver spacings as
+`dxes = [grid.autoshifted_dxyz(), grid.dxyz]`. The forward E curl needs
+distances between adjacent base-cell centers; the backward H curl needs
+distances between adjacent upper cell edges. Reversing these two width families
+introduces an error on graded grids that uniform grids hide.
+
+If constructing the mesh directly from **forward derivative intervals** `d`,
+with H samples at their midpoints, the backward intervals instead follow
+`(d + numpy.roll(d, 1)) / 2` for periodic boundaries. These `d` arrays are
+already center-to-center distances, whereas `grid.dxyz` contains base-cell
+volume widths. Applying this averaging rule to `grid.dxyz` does not fix the
+ordering error above.
+
 ::: meanas.fdmath
 
 ## Functional and sparse operators