Smaller Memory, Smaller Storage footprint and Faster Custom Dataset Creation

[bio-info-guy]

Modification to PertData for Smaller Memory, Smaller Storage footprint and Faster custom dataset creation

• use new GearsDataset class after data splitting to handle on the fly generation of data objects
• low_mem option to first store indices of x, y pairs instead of the actual x and y (no longer need to create pyg pickle objects, each time the dataset_processed is created directly either via loading the h5ad or creating new dataset)
• more filter steps of adata during dataset creation:
  • filter out singlets condition_names and contexts/cell_lines with no ctrl/ko pairings ( prevent some errors during DEG calculation on custom datasets)
  • filter adata during new_data_process to only include cells with perturbation in the gene GO graph.