FIX: correct transform_vectors for angular coordinates

[greglucas]

Rationale

Uses central differencing, falling back to single-sided forward or backward if necessary, and the local Jacobian via pyproj.get_factors() as suggested in the notes of #1926. This removes the need for special cutouts for angular projections and different branches and generalizes the computations for any projection <-> projection pair.

closes #1179

Implications

A few more transforms due to central differencing, but performance was not significantly worse in my local experimentation.

Alternative to #2718

[domfrbassett]

Hi @greglucas, I checked this against the #2279 linearity case locally, and the transform is linear in the input components as expected. The #1179 near-pole PlateCarree case also gives the expected direction.

One thing I noticed while testing: when self == src_proj, the new Jacobian path still runs even though the vector transform should be the identity. Adding an early return for that case preserves the existing result while avoiding unnecessary work, and it also means the test_barbs_1d.png baseline doesn't need to change.

I think this could just be:

if self == src_proj:
    return (np.array(u, dtype=np.float64, copy=True),
            np.array(v, dtype=np.float64, copy=True))

[greglucas]

Good idea @domfrbassett! Just added that cut-out for identical CRS'