Add Adaptive Computation Time (ACT) layer

[Cydral]

Implement adaptive_computation_time_ layer following Graves (2016) "Adaptive Computation Time for Recurrent Neural Networks".

This layer enables neural networks to adaptively determine the number of computation steps to perform for each sequence position, spending more computation on difficult parts while quickly processing easier parts.

Key features:

• Full CPU and CUDA implementations
• Configurable halting threshold and ponder penalty for regularization
• Support for depth-dependent gradient scaling
• Early termination optimization when all positions converge
• Template-based max_steps configuration (act4, act8, act16 aliases)

Note: This ACT implementation can serve as a building block for hierarchical reasoning mechanisms such as those described in "Hierarchical Reasoning Model" (Wang et al., 2025, arXiv:2506.21734) for advanced sequential reasoning tasks.

[davisking]

Nice, this is great. Thanks for the PR :D

[Cydral]

No problem. My next plan is to work on a native Transformer implementation to make them easier to use within Dlib.
In fact, I think I've brought almost everything needed: the existing template structure used in samples works well to maintain the library's style while allowing for flexible configuration (number of layers, vocab size, etc.). I'll also use this opportunity to add an image preprocessing layer, creating a unified framework for Transformers in Dlib.
After that, training a respectable-sized model will be the next challenge, of course... Perhaps, the place for another example to add ;-)