Fix MotionConv2d to cast blur_kernel to input dtype instead of reverse (#13364)

Co-authored-by: dg845 <58458699+dg845@users.noreply.github.com>

Author: YangKai0616

src/diffusers/models/transformers/transformer_wan_animate.py

@@ -166,8 +166,7 @@ def forward(self, x: torch.Tensor, channel_dim: int = 1) -> torch.Tensor:
             # NOTE: the original implementation uses a 2D upfirdn operation with the upsampling and downsampling rates
             # set to 1, which should be equivalent to a 2D convolution
             expanded_kernel = self.blur_kernel[None, None, :, :].expand(self.in_channels, 1, -1, -1)
-            x = x.to(expanded_kernel.dtype)
-            x = F.conv2d(x, expanded_kernel, padding=self.blur_padding, groups=self.in_channels)
+            x = F.conv2d(x, expanded_kernel.to(x.dtype), padding=self.blur_padding, groups=self.in_channels)
 
         # Main Conv2D with scaling
         x = x.to(self.weight.dtype)