feat(authoring): add stochastic-local-search terrain deformation (#409)

[barrulus]

Closes #409

Implements the constraint-based terrain deformation of Stachniak & Stuerzlinger, An Algorithm for Automated Fractal Terrain Deformation (WSCG 2005), as hmap::sls_deformation() in the authoring module.

What's in

• sls_deformation(): deforms a heightmap by a sequence of truncated-Gaussian pushes (location, amplitude, radius) chosen by stochastic local search to minimise a per-vertex fitness sum. One push per iteration; the best candidate is applied with probability p_best (0.65 as in the paper), otherwise one of the top_fraction best. The slope limit talus_max * radius implements the paper's frequency limitation.
• DeformationConstraint {target, weight, MATCH | ABOVE | BELOW, scale}: combinable fitness terms covering the paper's cases (match a heightmap, shape mask, fixed road, edge matching), see the header docs.
• GaussianPush + apply_gaussian_pushes(): the applied sequence can be recorded and replayed (paper §4.5).
• 11 gtests, example ex_sls_deformation (star island + S-shaped road at 512², a few seconds per run), docs/images/ex_sls_deformation.png.

Deviations from the paper

• The push amplitude is the least-squares optimum of the local quadratic penalty over the kernel footprint (exact for MATCH terms, active set for one-sided terms), clamped by the slope limit, instead of sampling 25 heights.
• Candidate vertices: half on a jittered uniform grid, half sampled proportionally to the current penalty map, instead of quad-tree refinement around features.
• The confidence-based prediction optimisation is not implemented.

CPU only (OpenMP over candidates); a GPU variant or applying several non-overlapping pushes per iteration would be natural follow-ups.