Fix: intersection_area & torch cpu device

1. The original function intersection_area was not correct.
Standard intersection x1 (left edge) is max(bbox_left[0], bbox_right[0]). Standard intersection y1 (top edge) is max(bbox_left[1], bbox_right[1]).
The current implementation uses min for both these calculations when determining the intersection dimensions (min(bbox_left[0], bbox_right[0]) and min(bbox_left[1], bbox_right[1])). This will not calculate the correct intersection rectangle dimensions. The intersection_area function [within remove_overlap_new function](https://github.com/microsoft/OmniParser/blob/5171b092483ab3e74ca50b9357e225f9f3571f18/util/utils.py#L242-L247) in the Omni github repo uses the correct standard formula (max for top-left, min for bottom-right).
2. An error occurs when using the CPU. This is because the pretrained model is loaded using torch16, while inference is only performed using torch16 on CUDA and MPS.

handler.py

@@ -296,7 +296,7 @@ class EndpointHandler:
                 return_tensors="pt",
                 do_resize=False,
             )
-            if self.device.type in {"cuda", "mps"}:
+            if self.device.type in {"cuda", "mps", "cpu"}:
                 inputs = inputs.to(device=self.device, dtype=torch.float16)
 
             with torch.inference_mode():
@@ -352,9 +352,14 @@ def area(bbox: List[int]) -> int:
 
 
 def intersection_area(bbox_left: List[int], bbox_right: List[int]) -> int:
-    return max(
-        0, min(bbox_left[2], bbox_right[2]) - min(bbox_left[0], bbox_right[0])
-    ) * max(0, min(bbox_left[3], bbox_right[3]) - min(bbox_left[1], bbox_right[1]))
+    ix1 = max(bbox_left[0], bbox_right[0])
+    iy1 = max(bbox_left[1], bbox_right[1])
+    ix2 = min(bbox_left[2], bbox_right[2])
+    iy2 = min(bbox_left[3], bbox_right[3])
+
+    width  = max(0, ix2 - ix1)
+    height = max(0, iy2 - iy1)
+    return width * height
 
 
 def intersection_over_union(bbox_left: List[int], bbox_right: List[int]) -> float: