avoid unused bias draws in linear no_bias

[Boulea7]

What this changes

Constructs the weight directly in Linear::no_bias, without generating a bias that is immediately discarded. This preserves the layer's weight initialization while removing the unused bias draws from the subsequent RNG stream.

Why

Fixes #51.

How it was verified

The seeded RNG-stream regression fails on unmodified master and passes with this change. The same tests also check ordinary biased initialization and bias-free forward/parameter behavior.

• cargo test --locked --offline -- --test-threads=1 passes (CPU tests and doctests)
• cargo clippy --locked --offline --all-targets -- -D warnings is clean
• cargo fmt --check passes

CUDA, Rust 1.87, and release-profile checks were not run locally.

Checklist

• Public items have /// docs

[CanReader]

Thanks, for your PR I checked that the new test fails on master and passes here, and CI is green currently.

There's one optional thing, not blocking: new and no_bias now both spell out the weight init. Conv2d handles this with a private build and bias.then(...), and the same works here (doc comments as before):

pub fn new(in_features: usize, out_features: usize) -> Linear {
    Linear::build(in_features, out_features, true)
}

pub fn no_bias(in_features: usize, out_features: usize) -> Linear {
    Linear::build(in_features, out_features, false)
}

fn build(in_features: usize, out_features: usize, bias: bool) -> Linear {
    Linear {
        weight: Param::new(Tensor::kaiming_uniform(
            &[out_features, in_features],
            in_features,
        )),
        bias: bias.then(|| Param::new(uniform_bias(out_features, in_features))),
        in_features,
        out_features,
    }
}

The weight is still drawn before the bias, so new_preserves_weight_then_bias_initialization passes unchanged. Happy to merge as is if you'd rather leave it.

[CanReader]

Ok I'm merging this PR now, for any improvments you can open a new PR, I'll review it