fix(tests): align test suite with refactored v2 API

[Copilot]

Tests were written against the pre-refactor API and fail across all Python versions (3.10–3.12) due to changed return types, renamed keys, and updated function signatures introduced in the v2 enhanced-analysis rewrite.

What changed

compute_efd — dict return instead of ndarray

# before
efd = compute_efd(contour, n_harmonics=10)
efd.shape  # AttributeError: 'dict' has no .shape

# after
efd["efd_coeffs"].shape  # (10, 4)

compute_radial_profile — renamed keys, removed radius_range_mu

• radius_* → radial_* (radius_mean_mu → radial_mean_mu, etc.)
• radius_range_mu dropped; range now derived as radial_max_mu − radial_min_mu
• n_radial_peaks added to expected key set

extract_features_from_array — signature and return type

• (mask, "name", resolution=4.25) → (mask, source_name="name", resolution_mu=4.25)
• Returns dict | None, not pd.DataFrame; assertions updated accordingly
• EFD key count corrected: 51 (40 coefficients + 10 power values + 1 deviation score)
• Hu moment column prefix: hu_ → hu (keys are hu1…hu7)

classify_shape — additional required keys

Added n_radial_peaks and asymmetry_index to all fixture feature dicts; the classifier now reads both.

curl_index_from_skeleton — nan result and wrong assertion direction

• _curved_skel was under-sampled (~120 pts for a 251-px arc), creating disconnected 1-px segments that all got skipped, yielding nan. Increased sample density to 4× arc pixel length.
• The function computes chord / arc (a straightness ratio, ≤ 1). A curved line therefore produces a lower value than a straight line. Assertion corrected to ci_curved < ci_straight.