Optimizations to LINGER

[arnavg115]

• GPU support: GPU support for training is added but, the performance increase is negligible to negative, likely because LINGER trains many smaller neural networks. The overhead of moving tensors etc. between the cpu and gpu outweighs the performance gains. Currently I defaulted to using the cpu since it's slightly faster.
• pd.concat: Moved pd.concat outside of loops. Due to how pd.concat copies dataframes it is inefficient to be run repetitively inside loops. Rather it's better to run once outside of a loop.
• Minor function improvements: I made minor improvements to a couple functions like cosine_similarity and introduced new versions that are slightly faster due to using torch under the hood.
• Vectorized T-tests: Rather than use an iterative variant, I introduced a vectorized variant which is slightly faster.
• Formatting: I ran the black python formatter and isort. This has no effect on functionality, but does make the code easier to read.