[WIP] Adds #[derive(Sequential)] on nn builder structs

[coreylowman]

Currently only tuples are supported for sequential networks. They are convenient and easy to write for small networks, but have a number of weaknesses when trying to scale up to larger networks.

1. The types get extremely long, making debug/error messages extremely difficult to read.
2. They are hard to work with when trying to accessing fields/subfields.

This PR introduces a new api for specifying sequential networks using a derive attribute:

#[derive(Sequential)]
struct MyMLP {
   layer1: Linear<5, 10>,
   act1: ReLU,
   layer2: Linear<10, 5>,
   act2: Tanh,
}

Internally, this will generate a new type that looks like this:

struct MyMLPBuilt<E: Dtype, D: Device<E>> {
   layer1: Linear<5, 10, E, D>,
   act1: ReLU,
   layer2: Linear<10, 5, E, D>,
   act2: Tanh,
}

and add all the TensorCollection/Module/ModuleMut implementations necessary.

This should be much easier to work with!

[coreylowman]

Noting that this effort has been moved into a separate repo https://github.com/coreylowman/slimnn. It's turned into a much bigger rewrite for the nn layer that will add:

1. #[derive(Sequential)] as mentioned in pr description
2. Simplify traits for nn layer (& remove TensorCollection)
3. Add derive macros for other new traits (ZeroGrads/ResetParams/UpdateParams/etc)
4. Add dynamically sized layers (Dynamic dimensions in neural network layers? #755).

I'm doing this work separately just for ease of implementation, and to see what it's like to implement a nn layer outside of the repo. Once that separate implementation is fairly complete, it will be merged into dfdx.