torchvision save_image produces incorrect results when saving png files using the mps backend

Issue description

torchvision save_image produces incorrect results when saving png files

in the attached images, you will see color pixels, but the input data is a rank two tensor so the images should be grayscale. please zoom in very far (800%) if you cannot see the red, yellow, etc color pixels.

Code example

for mnist (using mps as the device):

def test(model, device, test_loader):
    model.eval()
    test_loss = 0
    correct = 0
    with torch.no_grad():
        for data, target in test_loader:
            data, target = data.to(device), target.to(device)
            output = model(data)
            for i in range(0, 10):
               foo = data[i].reshape(28, 28).clone()
               torchvision.utils.save_image(foo, "foo" +str(i)+".png", normalize=True)

System Info

• PyTorch or Caffe2: pytorch
• How you installed PyTorch: conda
• Build command you used: not compiled from source
• OS: macOS Ventura 13.4
• PyTorch version: 2.0.1
• Python version: 3.9.16
• CUDA/cuDNN version: n/a (using mps backend)
• GPU models and configuration: M1 cpu (mps backend)

Start by reproducing the issue through torchvision.utils.save_image with the provided MNIST-style rank-two tensor, normalize=True, and the MPS device on macOS. Compare the generated PNG with CPU output and verify that the completed fix preserves grayscale pixels without color artifacts on MPS.