feat: kernel JIT batch top-max-per-variant (opt-in)

Nuovo kernel _jit_top_max_per_variant: prange esterno sulle varianti
invece di n_vars chiamate JIT separate via ThreadPoolExecutor.
Wrapper Python top_max_per_variant prepara buffer flat (offsets +
dx_flat/dy_flat/bins_flat) e bg per scala.

Default batch_top=False perche su benchmark realistici (Linux 13 core,
72-180 varianti) ThreadPoolExecutor + kernel singolo che rilascia GIL
e gia ottimale. Path batch_top=True utile come opzione per scenari
con n_vars >>> n_threads o overhead chiamate JIT dominante.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

pm2d/line_matcher.py

@@ -40,6 +40,7 @@ from pm2d._jit_kernels import (
     score_by_shift as _jit_score_by_shift,
     score_bitmap as _jit_score_bitmap,
     score_bitmap_rescored as _jit_score_bitmap_rescored,
+    top_max_per_variant as _jit_top_max_per_variant,
     popcount_density as _jit_popcount,
     HAS_NUMBA,
 )
@@ -574,6 +575,7 @@ class LineShapeMatcher:
         verify_threshold: float = 0.4,
         coarse_angle_factor: int = 2,
         scale_penalty: float = 0.0,
+        batch_top: bool = False,
     ) -> list[Match]:
         """
         scale_penalty: se > 0, riduce lo score per match a scala diversa da 1.0:
@@ -657,7 +659,25 @@ class LineShapeMatcher:
 
         kept_coarse: list[tuple[int, float]] = []
         all_top_scores: list[tuple[int, float]] = []
-        if self.n_threads > 1 and len(coarse_idx_list) > 1:
+        # batch_top: usa kernel batch single-call con prange-esterno su
+        # varianti. Vince su threadpool quando n_vars >> n_threads e quando
+        # H*W top e' piccolo (overhead chiamate JIT > costo kernel).
+        if (batch_top and HAS_NUMBA and len(coarse_idx_list) > 4):
+            dx_l = []; dy_l = []; bn_l = []; vs_l = []
+            for vi in coarse_idx_list:
+                var = self.variants[vi]
+                lvl = var.levels[min(top, len(var.levels) - 1)]
+                dx_l.append(lvl.dx); dy_l.append(lvl.dy); bn_l.append(lvl.bin)
+                vs_l.append(var.scale)
+            scores_arr = _jit_top_max_per_variant(
+                spread_top, dx_l, dy_l, bn_l, bg_cache_top, vs_l,
+                bit_active_top,
+            )
+            for vi, best in zip(coarse_idx_list, scores_arr.tolist()):
+                all_top_scores.append((vi, best))
+                if best >= top_thresh:
+                    kept_coarse.append((vi, best))
+        elif self.n_threads > 1 and len(coarse_idx_list) > 1:
             with ThreadPoolExecutor(max_workers=self.n_threads) as ex:
                 for vi, best in ex.map(_top_score, coarse_idx_list):
                     all_top_scores.append((vi, best))