Describe the bug
When the metatensor has pending operations, modifying the applied_operations property may change the final transform output's applied_operations ordering.

for example, the following LambdaD pushes to the applied_operations while pending_operations is trying to combine multiple lazy operations:

Compose(
    [
        ResizeWithPadOrCropD(keys=["img", "seg"], spatial_size=[80, 72, 80]),
        Rotated(keys=["img", "seg"], angle=[np.pi / 2, np.pi / 2, 0], mode="nearest", keep_size=False),
        Lambdad(keys=["img"], func=lambda x: torch.sigmoid((x - x.min()) / x.std())),
    ],
    lazy_evaluation=True,
)

this becomes an issue when we try to inverse the applied_operations.