Scale the DLR round-trip tolerance in the axis test by its conditioning

[shinaoka]

Companion to SpM-lab/sparse-ir-rs#326.

test_axis_on_three_dimensional_input compares the same logical data with the
coefficient axis in two places. The transform applies the same linear map either
way, but the contraction geometry can differ, so the two results differ by their
rounding amplified by the transform's conditioning. The three sampling
transforms are well conditioned and keep the fixed 1e-12 bound; the DLR round
trip is not (cond(-S_l V_l(w_p)) is ~2.2e5 for the fixture here), and a library
built from sparse-ir-rs main misses that bound by 2% (absolute 3.24e-12
against atol 3.19e-12).

The tolerance is now derived from the fixture's own conditioning rather than
hard-coded:

amp = np.linalg.cond(-basis.s[:, None] * basis.v(obj.sampling_points))
atol = max(1e-12, amp * np.finfo(np.float64).eps) * np.abs(back_ref).max()

Test only: no runtime code changes, and nothing to release.

Verified:

• pytest tests/test_boundary.py -k axis_on_three -> 8 passed, both against the
  published pylibsparseir 0.10.0 wheel and against a library built from
  sparse-ir-rs main (4 failed before).
• Full suite -> 356 passed against both libraries.
• check_libsparseir_version_consistency.py -> all checks passed.