Fixing the inefficiencies & loopholes in previous version

[shivendrra]

To be done:

Restructuring

• Restructuring entire project (fixing some components here & there)

axon/
├── csrc/    // all c/cpp, cuda codes in here
│   ├── cpu/    // all cpu BLAS, SIMD & OMP related kernels
│   ├── cuda/    // all cuda kernels
│   ├── inc/     // random.h
├── nn/    // all neural net components like 'module', 'loss', 'optim'
│   ├── modules/   // 'linear', 'embedding', 'conv2d'
├── linalg/  // python version of linalg functions
│   ├── __init__.py
│   ├── base.py   // interface for python-cpp (compiled) code
│   ├── tensor.py   // main tensor class lies here
│   ├── helpers.py   // some python helper functions for tensor class
│   ├── utils.py    // 'ones_like', 'zeros_like', etc
├── build/
├── tests/
├── CMakeLists.txt
└── run.py

Fixing Inefficiencies

• Re-writing the core array class with dtype support
• Device interface to switch b/w cpu & gpu
• Fixing axon.linalg functions
• Update test cases
• Update & re-write documentation
• Python inefficiencies removed, fixing the interfacing delay