Lightning 2.5.5 breaks checkpoint reuse

[ziw-liu]

We have workflows where checkpoints trained with one wrapper class is reused for another. For example, a FullyConvolutionalMAE wrapped in FcmaeUNet during training could be wrapped in AugmentedPredictionVSUNet for inference. Lightning-AI/pytorch-lightning#21116 shipped in lightning 2.5.5 breaks this use case as it raises an exception if the hyperparameter does not match what's saved in the checkpoint.