Initial graph

[vadixidav]

This creates a graph that is not ideal (see issue #4 for more details), but does work for most machine learning use cases anyways. The approach in the PR initially is to represent all of the available ops that all backends can execute using an enum called Op. This approach is likely not what we ultimately intend to use since it doesn't support adding ops downstream, but this PR will be open to discuss this.

The deep crate contains the core graph logic and tensor logic. It also contains what defines a backend.

The deep-native crate defines the backend for native Rust implemented operations. It specifically chooses to use ndarray to represent its tensors. It allows ops to be plugged into it at runtime.

The deep-backend-tools crate was created because deep-native had several pieces of code that likely apply to all backends. This crate is intended to make the development of future backends easier by allowing the reuse of code. This crate performs the graph forward and backward propogation using a Tape struct that remembers previous outputs to avoid re-computation when gradients need to be applied.