Port test_p2e2p to use expansion toys

[Copilot]

test_p2e2p was written against the low-level P2EFromSingleBox/E2PFromSingleBox/P2P API with manual box index arrays. The expansion toys module (sumpy.toys) provides a cleaner abstraction for exactly this kind of test, as already used in test_translations and test_toy_p2e2e2p.

Changes

• Replaced low-level kernel calls with t.ToyContext, t.PointSources, t.multipole_expand/t.local_expand, and .eval(). All manual box management arrays (source_boxes, box_source_starts, box_source_counts_nonchild, etc.) are removed — handled internally by the toys layer.

• Split extra_kwargs → extra_source_kwargs + extra_kernel_kwargs as required by ToyContext (kernel parameters like k need to reach both P2E and E2P).

• Moved expansion construction outside the h-loop — the center and sources are fixed across h-values; only evaluation targets change. The expansion is computed once and reused.

• Dropped the AxisTargetDerivative gradient test — target derivatives require a separate kernel evaluation and don't fit the one-kernel-at-a-time toys API. The h-convergence check for the primary potential is preserved in full.

• Removed unused obj_array import (was only needed for actx.from_numpy(obj_array.new_1d(fp.points)), now handled inside _e2p).

# Before: ~80 lines of box arrays, P2EFromSingleBox, E2PFromSingleBox, P2P, actx conversions

# After:
toy_ctx = t.ToyContext(knl, local_expn_class=expn_class,
                       extra_source_kwargs=extra_source_kwargs,
                       extra_kernel_kwargs=extra_kernel_kwargs)
point_sources = t.PointSources(toy_ctx, sources, strengths)
expn = t.local_expand(actx, point_sources, loc_center, order=order, rscale=rscale)

for h in h_values:
    fp = FieldPlotter(loc_center, extent=h, npoints=res)
    pot = expn.eval(actx, fp.points)
    pot_direct = point_sources.eval(actx, fp.points)
    eoc_rec_pot.add_data_point(h, la.norm(pot - pot_direct) / la.norm(pot_direct))