monai.transforms.intensity.array.NormalizeIntensity bug when __call__ with channel_wise=True

[Elferodie]

Describe the bug
In the __call__ method of monai.transforms.intensity.array.NormalizeIntensity class, the conversion into torch.float32 performed in self._normalize line 911, is not taken into account when channel_wise = True. The following code does not define a new variable img with the right type but it keeps the type of the original img variable. Therefore, if the original image was filled with int, the normalization will be performed on int and the normalization will fail.

for i, d in enumerate(img):
    img[i] = self._normalize(  # type: ignore
        d,
        sub=self.subtrahend[i] if self.subtrahend is not None else None,
        div=self.divisor[i] if self.divisor is not None else None,
    )

To Reproduce
Steps to reproduce the behavior:
Run commands

from monai.transforms.intensity.array import NormalizeIntensity
tensor = torch.arange(24).reshape(2,3,4
normalized_correctly = NormalizeIntensity()(tensor)
normalized_corrupted = NormalizeIntensity(channel_wise=True)(tensor)
print(normalized_correctly)
print(normalized_corrupted)

Expected behavior

>>> normalized_correctly
metatensor([[[-1.6613, -1.5169, -1.3724, -1.2279],
         [-1.0835, -0.9390, -0.7945, -0.6501],
         [-0.5056, -0.3612, -0.2167, -0.0722]],

        [[ 0.0722,  0.2167,  0.3612,  0.5056],
         [ 0.6501,  0.7945,  0.9390,  1.0835],
         [ 1.2279,  1.3724,  1.5169,  1.6613]]])
>>> normalized_corrupted
metatensor([[[-1., -1., -1.,  0.],
         [ 0.,  0.,  0.,  0.],
         [ 0.,  1.,  1.,  1.]],

        [[-1., -1., -1.,  0.],
         [ 0.,  0.,  0.,  0.],
         [ 0.,  1.,  1.,  1.]]])