Method linear_forward

cake-core/src/backends/mod.rs:206–241

Linear layer forward: `x @ weight^T + bias`. Matches candle_nn::Linear::forward() semantics exactly: - For contiguous 3D/4D inputs: reshape to 2D → matmul → reshape back (avoids slow broadcast_matmul on CUDA/CPU) - For non-contiguous 3D+: uses broadcast_left on weight - No dtype conversion (caller is responsible)

(
        &self,
        x: &Tensor,
        weight: &Tensor,
        bias: Option<&Tensor>,
    )

    /// - For non-contiguous 3D+: uses broadcast_left on weight
    /// - No dtype conversion (caller is responsible)
    fn linear_forward(
        &self,
        x: &Tensor,
        weight: &Tensor,
        bias: Option<&Tensor>,
    ) -> Result<Tensor> {
        let out = match x.dims() {
            [b1, b2, m, k] => {
                if x.is_contiguous() {
                    let w = weight.t()?;
                    x.reshape((b1 * b2 * m, *k))?
                        .matmul(&w)?
                        .reshape((*b1, *b2, *m, ()))?
                } else {
                    let w = weight.broadcast_left((*b1, *b2))?.t()?;
                    x.matmul(&w)?
                }
            }
            [bsize, m, k] => {
                if x.is_contiguous() {
                    let w = weight.t()?;
                    x.reshape((bsize * m, *k))?
                        .matmul(&w)?
                        .reshape((*bsize, *m, ()))?
                } else {
                    let w = weight.broadcast_left(*bsize)?.t()?;
                    x.matmul(&w)?
                }
            }
            _ => x.matmul(&weight.t()?)?,
        };
        match bias {
            Some(b) => out.broadcast_add(b),
            None => Ok(out),
        }
    }

    /// RMS normalization: `x * weight / sqrt(mean(x^2) + eps)`.
    fn rms_norm(&self, x: &Tensor, weight: &Tensor, eps: f32) -> Result<Tensor> {