Make the Affine 32 Again

[bill-yc-chen]

Is your feature request related to a problem? Please describe.
When read in the NiftiImage with LoadImage transform the affine matrix is saved in torch.float64, which causes problems when using MacBook's neural engine because mps accelerator doesn't support float64.

Describe the solution you'd like
Just make the affine matrix into float 32, thank you!

Describe alternatives you've considered
None

Additional context
None

[KumoLiu]

Looks like a similar issue with #5511 and #3831.
Could you please try monai.transforms.CastToType first as a workaround?

[wyli]

also I think you can bypass the setter with meta_tensor.meta['affine'] = mat4x4 instead of meta_tensor.affine = mat4x4

the current implementation is hard-coded to float64

MONAI/monai/data/meta_tensor.py

Line 470 in ef7debe

self.meta[MetaKeys.AFFINE] = torch.as_tensor(d, device=torch.device("cpu"), dtype=torch.float64)

otherwise there's some precision issues when using it for resampling.