Save individual normalization modes (multi-channel) in Auto3dSeg's model config

Is your feature request related to a problem? Please describe.

I trained a model with the Auto3dSeg Autorunner and was trying to run inference using the SlicerMONAIAuto3DSeg extension (https://github.com/lassoan/SlicerMONAIAuto3DSeg).

modality: mri
extra_modalities: {image2 : none}    # a second modality is a label which should NOT be normalized

The model expects multiple input images (image1: mri, image2: label --> no normalization).

The following .yaml file is an example of the use of multiple input images for a model: https://github.com/Project-MONAI/tutorials/blob/main/auto3dseg/tasks/hecktor22/input.yaml

Currently, no matter how many input modalities (imageModality and extra_modalities) were provided, only a single value for normalize_mode is saved in the model's `config'.

Describe the solution you'd like

1. When multiple input images were provided and the same normalization mode was used for all of them, then a single value fornormalize_mode is valid
2. When multiple input images of different types/modalities were provided, then a list of normalization modes corresponding to the number of input images should be used for normalize_mode.

Additional context

https://github.com/lassoan/SlicerMONAIAuto3DSeg/issues/32
https://github.com/JolleyLab/Internal/issues/194

Start with the Auto3dSeg Autorunner path that writes the model config and compare its handling with the tutorials' auto3dseg/tasks/hecktor22/input.yaml example. Trace how imageModality and extra_modalities become normalize_mode values. Done means identical modes remain a single value, while differing modes are saved in input order as a list, including an unnormalized label modality.