perf: enable torch.compile for SAM2 via vos_optimized flag

Uses Facebook's official SAM2VideoPredictorVOS which compiles all five
model components (image_encoder, memory_encoder, memory_attention,
sam_prompt_encoder, sam_mask_decoder) with torch.compile max-autotune.
First inference has ~30s warmup cost; subsequent frames benefit from
fused kernels and reduced memory round-trips.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

models/segmenters/grounded_sam2.py

@@ -362,8 +362,11 @@ class GroundedSAM2Segmenter(Segmenter):
         from sam2.sam2_image_predictor import SAM2ImagePredictor
 
         # Video predictor (for process_video)
+        # vos_optimized=True enables SAM2VideoPredictorVOS which compiles
+        # image_encoder, memory_encoder, memory_attention, sam_prompt_encoder,
+        # and sam_mask_decoder with torch.compile(mode="max-autotune").
         self._video_predictor = build_sam2_video_predictor_hf(
-            hf_id, device=self.device
+            hf_id, device=self.device, vos_optimized=True,
         )
 
         # Image predictor (for single-frame predict)