fix: use ImageNet normalization in latent_feature.ipynb (#53)

[parallelArchitect]

Fixes #53

latent_feature.ipynb was normalizing each image channel by its own per-image mean and std:

img[...,0] = (img[...,0] - img[...,0].mean())/img[...,0].std()

This is inconsistent with the training pipeline in util/datasets.py which uses fixed ImageNet statistics:

from timm.data.constants import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
transforms.Normalize(mean, std)

Fixed by replacing per-image normalization with the same ImageNet constants used during training:

imagenet_mean = np.array([0.485, 0.456, 0.406])
imagenet_std  = np.array([0.229, 0.224, 0.225])
img = (img - imagenet_mean) / imagenet_std

numpy is already imported in the notebook so no new dependencies are required.