Update yolo.py

Checking for `onnx_dynamic` first should suppress the warning:

```log
TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if self.grid[i].shape[2:4] != x[i].shape[2:4] or self.onnx_dynamic
```

Author: deepsworld

models/yolo.py

@@ -55,7 +55,7 @@ def forward(self, x):
             x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
 
             if not self.training:  # inference
-                if self.grid[i].shape[2:4] != x[i].shape[2:4] or self.onnx_dynamic:
+                if self.onnx_dynamic or self.grid[i].shape[2:4] != x[i].shape[2:4]:
                     self.grid[i], self.anchor_grid[i] = self._make_grid(nx, ny, i)
 
                 y = x[i].sigmoid()