Add linear_ layer for neural networks

[Cydral]

This PR adds a new linear_ layer that provides an optimized linear transformation for multi-dimensional inputs. The layer preserves the spatial dimensions of the input tensor while transforming the feature dimension.

Features

• Supports both biased and unbiased linear transformations
• Optimized for maintaining tensor dimensions (except for feature dimension)
• Compatible with existing DNN API patterns and serialization

This implementation is more efficient than reshaping inputs to use fc_ layers when working with multi-dimensional data.
Similar functionality exists in other major frameworks like PyTorch (nn.Linear), which operates on the last dimension while preserving other dimensions.

[Copilot]

Pull Request Overview

This PR introduces a new linear_ layer that provides an optimized linear transformation for multi-dimensional inputs while preserving spatial dimensions except for the feature one. The implementation supports both biased and unbiased modes and is integrated into the DNN API with corresponding test cases added.

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
dlib/test/dnn.cpp	Added multiple tests exercising linear_ layer instantiations (e.g. linear_<1, LINEAR_NO_BIAS>) and a dedicated test_linear() function to validate outputs and gradients.
dlib/dnn/layers_abstract.h	Introduced the abstract definition for the new linear_ layer and the associated linear_bias_mode enum with clear documentation.
dlib/dnn/layers.h	Provided concrete implementations of the linear_ layer, including input/output mapping and wrapper aliases (linear and linear_no_bias).

Comments suppressed due to low confidence (1)

dlib/test/dnn.cpp:3561

• [nitpick] Consider renaming 'xtmp' to a more descriptive name (like 'input_matrix') to improve code clarity in the test_linear() function.

matrix<float> xtmp(2, 4);

[davisking]

Nice, thanks for the PR 😁