KNN computation using Keops post dataloader

[chaidosa]

In this branch, the key change is the introduction of post_knn in the config file, allowing KNN computation on the GPU after the dataloader instead of within the CPU-based dataloader. We use KeOps for KNN computation post-dataloader.

The file knn_post_dataloader_train.py provides an example of how to implement this, while knn_post_dataloader_utils.py contains the necessary functions for retrieving correct values and performing KNN computation using KeOps.

Additionally, the previous functionality remains unchanged—if post_knn is not specified in the config file, the codebase functions as in the previous version.

[chaidosa]

@npmhung, I have made the changes as we discussed, updated the README, renamed the files for better readability, and included the config file for post_knn usage.

[npmhung]

what does "prepare" function do exactly?

[chaidosa]

@npmhung, the prepare function is used to create a packed representation of the kNN indices. it is similar to datasetCommon.py which is used in the dataloader.