ADD embedding based sampling example

[future-158]

This pull request introduces an embedding-based sampling method in an active learning pipeline for CIFAR-10. The key update is the way we select informative samples from the pool dataset after each model retraining cycle.

Embedding-Based Sampling Method

After retraining the model, we evaluate the validation set to calculate the 'hardness' of each image.
Hardness is determined based on the model's prediction confidence for the true labels, calculated as 1 - probability of the true class.
We then use the embeddings generated by the model to match each image in the pool dataset with a corresponding image in the validation set.
The pool dataset images are assigned a 'hardness' score indirectly, based on their closest match in the validation set's embedding space.
This method allows us to prioritize learning from images in the pool that are similar to those the model finds challenging.