Fix accuracy-order loss for even derivatives on non-uniform grids

[gaoflow]

On a non-uniform grid the interior central stencil for every even-order derivative is sized as if it earned the uniform-grid symmetry bonus (even-derivative central differences gain one order for free on a symmetric grid). That bonus does not exist on a non-uniform grid, so the central stencil ends up one point too small — one fewer point than findiff's own forward/backward stencils — and the interior accuracy drops from acc to acc-1. The non-uniform Laplacian is therefore only first-order accurate in the interior. Odd-order derivatives are unaffected.

The sizing lives in coefs.py in both non-uniform code paths (coefficients_non_uni and calc_coefs_non_uni_batched), so it propagates through both Diff.__call__ and the .matrix() operator path (PDEs, eigenvalue problems). The fix sizes the central stencil like the one-sided schemes, num_coef // 2 points per side, so it reaches order acc without relying on the symmetry bonus. For odd derivatives num_coef // 2 equals the previous value, so those paths are unchanged; the matrix/apply machinery is data-driven off the returned offsets and num_bndry, so no further changes are needed.

Verified three ways:

• Polynomial exactness: an order-acc scheme reproduces polynomials of degree deriv+acc-1 exactly. On a graded grid the even-derivative interior was off by ~1e-2 before and is now at machine precision.
• Graded-grid convergence order: d2/dx2 at acc=2 measured order 1.0 before (want 2), acc=4 order 3.0; both now reach their requested order in the interior and in the .matrix() operator.
• Fornberg-weight cross-check: the computed weights already matched Fornberg (1988) to ~1e-11 for the chosen stencil — the weights were correct, only the stencil width was wrong; adding one point cut the pointwise error from 3e-5 to 3e-9.

Regression tests in tests/test_nonuniform_accuracy.py cover polynomial exactness and convergence order for even and odd derivatives across accuracy orders, with uniform and odd-derivative controls. Three existing tests asserted the undersized sizing directly (that the even-derivative central stencil equals the equidistant/one-sided one on a uniform grid); they are updated to the corrected sizing. Full suite passes.