feat: scene precompute cache (II Halcon-style)
LRU cache per scena: hash su prime 64KB bytes + parametri matcher
(weak/strong_grad, spread_radius, n_bins, pyramid_levels). Quando
hit, riusa:
- piramide grays
- spread_top + bit_active_top + density_top
- spread0 + bit_active_full + density_full
Tipico use case: UI tuning con slider min_score/verify_threshold/...
produce 10+ find() consecutive su scena identica. Risparmia
Sobel+dilate+popcount duplicati (~50ms su 1080p).
Speedup misurato: ~15% find() su 1080p (54ms su 351ms). Vantaggio
maggiore su template piccoli (kernel JIT veloce → scena precompute
domina). Cache size 4, invalidata in train() (template cambiato).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adriano
74a332a2dd