Add tril_ layer for lower triangular matrix operations

[Cydral]

This PR introduces a new tril_ layer to dlib, which implements lower triangular matrix operations similar to PyTorch's torch.tril() function. The layer allows for flexible lower triangular matrix masking with customizable diagonal offset and diagonal value.

Key features:

• Implements lower triangular masking for tensors
• Supports custom diagonal offset and diagonal value
• Offers three convenient alias templates: tril, tril_mask, and tril_diag

This addition enhances dlib's neural network capabilities, allowing for more complex architectures that require lower triangular matrix operations.
The new layer can be particularly useful in attention mechanisms, triangular matrix operations, and other scenarios where lower triangular masking is required in neural network architectures.