[WIP] Option for using GPU

[pskrunner14]

• Adds Numba functional API for offloading computation to CUDA capable device.
• Adds CUDA C interface (still buggy so defaults to numba).
• Refactored core layers API.
• Using autograd for loss derivatives.

note: fixes #2

[pskrunner14]

TODO:

• Resolve merge conflicts.
• Use cuBLAS for matrix ops and benchmark performance against OpenAAC and OpenMP for larger inputs and hidden units.
• Fix numerical gradient checking and improve tests for CI consistency.
• Refactor code for future maintainability.