fix: use circular mean for collocated dip direction aggregation

[lachlangrose]

Summary

• DipDipDirectionInterpolator.setup_interpolation clusters structure points that are within cell_size of each other (DBSCAN) and averages X, Y, DIP and DIPDIR within each cluster with a plain .mean().
• DIPDIR is a compass bearing (wraps at 360°). A plain arithmetic mean is wrong whenever a cluster straddles due north: averaging 350° and 10° gives 180° — the exact opposite direction — instead of the correct 0°.
• DIPDIR is now aggregated with a circular mean (mean of unit vectors, converted back to an angle). X, Y and DIP are ordinary numeric quantities and are unaffected.

Note: this fixes aggregation of collocated points only. The RBF interpolation step across the full grid still blends raw DIPDIR values without circular awareness, so it can still produce artifacts near a 0°/360° discontinuity — that's a separate, larger change (e.g. interpolating sin/cos components independently) and out of scope here.

Test plan

• pytest tests/thickness tests/project/test_thickness_calculations.py — passes unchanged (exercises InterpolatedStructure, which uses this interpolator)
• Manually verified _circular_mean_degrees([350, 10]) == 0 (mod float precision) vs the old mean([350, 10]) == 180

[lachlangrose]

Closing — map2loop now lives in packages/map2loop of Loop3D/LoopStructural, and this repository is being archived. Ported to Loop3D/LoopStructural#328