INFERENCE-TIME IO NORMALIZATION: Handling images with a number of channels different to any in the available pixel statistics.

[AlexSath]

Description:

Similarly to #6, if an inference-time input image contains a number of channels that does not match the size of any 1D array in the 'FLAME_Dataset' json, current behavior lead to an error and a crash. Need to decide how to handle such a mismatch in the future.

Options:

1. Logger mismatch as warning, skip for inference
2. Min-max-scale using percentiles calculated from the pixel distribution of the singular image
3. Allow the user to override the percentile values for min-max scaling
4. Aggregate available percentile information in 'FLAME_Dataset' json into a single min and max percentile for normalization. An example could be the median of available channel statistics being applied to every channel in the input image with a mismatched number of channels.