Allow to train dinov2 with different dtypes like bf16 (#28504)

I want to train dinov2 with bf16 but I get the following error in https://github.com/huggingface/transformers/blob/bc72b4e2cdcbc80d5f56731f35dbc9c18b4c8de6/src/transformers/models/dinov2/modeling_dinov2.py#L635:

```
RuntimeError: Input type (float) and bias type (c10::BFloat16) should be the same
```

Since the input dtype is torch.float32, the parameter dtype has to be torch.float32...

@LZHgrla and I checked the code of clip vision encoder and found there is an automatic dtype transformation (https://github.com/huggingface/transformers/blob/bc72b4e2cdcbc80d5f56731f35dbc9c18b4c8de6/src/transformers/models/clip/modeling_clip.py#L181-L182).

So I add similar automatic dtype transformation to modeling_dinov2.py.

src/transformers/models/dinov2/modeling_dinov2.py

@@ -103,12 +103,13 @@ class Dinov2Embeddings(nn.Module):
         height, width = height + 0.1, width + 0.1
         patch_pos_embed = patch_pos_embed.reshape(1, int(math.sqrt(num_positions)), int(math.sqrt(num_positions)), dim)
         patch_pos_embed = patch_pos_embed.permute(0, 3, 1, 2)
+        target_dtype = patch_pos_embed.dtype
         patch_pos_embed = nn.functional.interpolate(
-            patch_pos_embed,
+            patch_pos_embed.to(dtype=torch.float32),
             scale_factor=(float(height / math.sqrt(num_positions)), float(width / math.sqrt(num_positions))),
             mode="bicubic",
             align_corners=False,
-        )
+        ).to(dtype=target_dtype)
         if int(height) != patch_pos_embed.shape[-2] or int(width) != patch_pos_embed.shape[-1]:
             raise ValueError("Width or height does not match with the interpolated position embeddings")
         patch_pos_embed = patch_pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
@@ -116,7 +117,8 @@ class Dinov2Embeddings(nn.Module):
 
     def forward(self, pixel_values: torch.Tensor, bool_masked_pos: Optional[torch.Tensor] = None) -> torch.Tensor:
         batch_size, _, height, width = pixel_values.shape
-        embeddings = self.patch_embeddings(pixel_values)
+        target_dtype = self.patch_embeddings.projection.weight.dtype
+        embeddings = self.patch_embeddings(pixel_values.to(dtype=target_dtype))
 
         if bool_masked_pos is not None:
             embeddings = torch.where(