Optionally use Vec for input storage

[HyperCodec]

The current const generics make it hard for external tools to deserialize a network that could be of any given type. Thus,I propose the following:

trait Dims {
    // might be able to simplify this with AsRef<[Neuron]> and AsMut<[Neuron> or the Deref traits
    type InputLayer
    where
        Self::InputLayer: Index<usize, Output = Neuron> + IndexMut<usize>,
        for<'a> &'a Self::InputLayer: IntoIterator<Item = &'a Neuron>,
        for<'a> &'a mut Self::InputLayer: IntoIterator<Item = &'a mut Neuron>;
    type OutputLayer; // ... same constraints

    fn init_input_layer(&self) -> Self::InputLayer;
    fn init_output_layer(&self) -> Self::OutputLayer;
}

struct NeuralNetwork<D: Dims> { ... }

impl<D: Dims> NeuralNetwork<D> {
    pub fn new(dims: &D, ...) -> Self {
        // ... gets empty collections via init_input_layer and init_output_layer
    }
}

#[derive(Debug, Copy, Clone, Default ...)]
struct StaticDims<const I: usize, const O: usize>; // uses fixed-length arrays

#[derive(Debug, Clone, ...)]
struct DynamicDims {
    input: usize,
    output: usize,
}; // uses Vec

StaticDims and DynamicDims should be compatible with each other for serialization as well, meaning a neural network could be trained with fixed-length optimizations and then easily interpreted by a tool which does not know its input and output lengths at compile time.