This PR adds webgpu support for neuralNetwork so it can be used without called ml5.setBackend.

Changes:

• bump TensorFlow.js dependencies to a version that supports readSync as @lindapaiste suggested.
• add await tf.ready() to the init function of DiyNeuralNetwork @lindapaiste suggested.
• change copy and crossover to async, support both promise and callback interface.
• Update example sketches accordingly

The ml5.neuralNetwork function is technically an async operation like with other models. It should be called with a callback function or placed within p5 preload function. Our examples, however, have been treating ml5.neuralNetwork like a synchronous function since it sometimes (when not loading data) behaves synchronously. The changes from this PR adds await tf.ready() thus making it definitively asynchronous.

The functions copy and crossover also instantiate the neuralNetwork object, which is why they have to be async.

I updated all example sketches, removing ml5.setBackend and moving the ml5.nerualNetwork function to preload for non neuroevolution examples.

The neuroevolution examples are kept on cpu backend for performance. Adapting for async copy and crossover slightly complicates those examples but I think they are still fairly readable. The ml5.neuralNetwork calls in neuroevolution examples are still synchronous, but they seem to not cause any error for now. I will try to come up with a clean way to adapt them to async.