[ENH] Improve performance by 2000% by directly batching without using dataloader

[jobs-git]

Is your feature request related to a problem? Please describe.
Data Loader from pytorch was designed for accessing files from Disk, however, this results to a very slow and very low utilization of high performance accelerators resulting to slower completion and inefficient training.

This can be alleviated by manually batching via for loop, and can result to about 2000% speed up! So that is very much performance improvement.

Describe the solution you'd like
De-couple the pytorch from pytorch lighting to enable manual batching

Describe alternatives you've considered
n/a

Additional context
n/a