fix(codec): merge per-canvas image_grid_thw into [[N,H,W]] per video

The codec branch in __call__ was emitting per-canvas image_grid_thw rows
(N, 3) = [[1, h, w]]*N, which caused the vision encoder's _build_cu_seqlens
(fixed_t=4) to treat each canvas as an isolated 1-frame sample instead of
grouping canvases into 4-frame attention windows. This silently degraded
video benchmark scores by ~1pt on videoeval_pro (45.07 vs reference 46.4).

Merge same-size canvases per video into a single [[N, H, W]] row so the
encoder restores the cross-canvas self-attention windows that the training
pipeline assumes.

processing_llava_onevision2.py

@@ -308,7 +308,30 @@ class LlavaOnevision2Processor:
                 all_patch_positions.append(patch_positions)
 
             out["pixel_values"] = torch.cat(all_pixel_values, dim=0)
-            out["image_grid_thw"] = torch.cat(all_grid_thw, dim=0)
+            # Merge per-canvas grid_thw rows into a single [N, H, W] row per
+            # video so the vision encoder's _build_cu_seqlens (fixed_t=4)
+            # groups canvases into 4-frame attention windows, matching the
+            # training pipeline. Without this merge the encoder treats each
+            # canvas as an isolated 1-frame sample (no cross-canvas
+            # self-attention), which silently degrades video benchmark
+            # scores by ~1pt on videoeval_pro.
+            merged_grid_thw_rows = []
+            for grid in all_grid_thw:
+                if (
+                    grid.shape[0] > 1
+                    and bool(torch.all(grid[:, 1] == grid[0, 1]).item())
+                    and bool(torch.all(grid[:, 2] == grid[0, 2]).item())
+                ):
+                    merged_grid_thw_rows.append(
+                        torch.tensor(
+                            [[int(grid.shape[0]), int(grid[0, 1]), int(grid[0, 2])]],
+                            dtype=grid.dtype,
+                            device=grid.device,
+                        )
+                    )
+                else:
+                    merged_grid_thw_rows.append(grid)
+            out["image_grid_thw"] = torch.cat(merged_grid_thw_rows, dim=0)
             out["patch_positions"] = torch.cat(all_patch_positions, dim=0)
             text = rewritten_texts
             # Codec branch handled the video. Suppress the frame-sampling block below.