Add LinearLayer for output projections and fully connected layers

[kholdrex]

The library currently lacks a dedicated linear/dense layer, which is essential for:

• Text generation (LSTM hidden, vocabulary logits)
• Classification tasks (hidden, class probabilities)
• Custom output projections

Proposed Implementation:

pub struct LinearLayer {
    weight: Array2<f64>,     // (output_size, input_size)
    bias: Array1<f64>,       // (output_size,)
}

impl LinearLayer {
    pub fn new(input_size: usize, output_size: usize) -> Self
    pub fn forward(&self, input: &Array2<f64>) -> Array2<f64>
    pub fn backward(&self, grad_output: &Array2<f64>) -> (LinearGradients, Array2<f64>)
}

Benefits:

• Enables standard text generation architectures
• Simplifies classification tasks
• Reduces boilerplate code for users