Fix CUDA build on Rust 1.77

[zeux]

• CUDA Compiler trait impls didn't define Output type so --features cuda wouldn't compile
• nn module wasn't exporting submodules as public so imports like this in Mistral demo didn't resolve:

use luminal::{
    nn::{embedding::Embedding, norm::RMSNorm},

After these fixes, cargo run --release --features cuda works from examples/mistral on Rust 1.77.

[jafioti]

@zeux Looks good! Only nitpick I have is changing all the nn modules to pub. Can you instead just change this:

use luminal::{
    nn::{embedding::Embedding, norm::RMSNorm},

to

use luminal::prelude::*;

because all the nn modules are re-exported by nn, which should be covered by prelude.

[zeux]

Yeah definitely, I can fix the import side - wasn't sure which direction to go.

[jafioti]

Looks great, thanks!