feat: Allow partial execution of network

[retraigo]

This PR adds the ability to execute a partial sequence of the network. This can be useful when designing models such as Autoencoders and GANs.

• Allow partial execution of the network.
• Add proper error handling for misconfiguration during partial execution (out-of-bound slicing, etc),
• Write an example showcasing how to execute the network partially (Autoencoders?)

And...

• Additionally added the ability to predict multiple input points in a single net.predict() call.

This is a breaking change for the predict() method. The following things are broken:

• Existing Examples
• GPU backend
• WASM backend

This PR also fixes #38.

EDIT:

• Rewrite every example to use new predict() format.
  • xor (all) (edit: skipping xor.ipynb for now)
  • linear
  • multiple-linear
  • binary
  • multiclass
  • text
  • filters
  • run
  • mnist-predict

[retraigo]

On the flip side, the last change I mentioned now requires rewriting every example to follow the new predict() format (eg. Using a tensor2D for XOR instead of tensor1D).

[retraigo]

I'm updating just the predict routine of the mnist example without disturbing the rest for now.

[retraigo]

The last commit fixes #38

[load1n9]

LGTM