Adding attention strategy

[moderouin]

This PR introduces two new strategies based on transformer-style neural networks with attention mechanisms:

1. Attention: A base strategy with a randomly initialized neural network
2. EvolvedAttention: A strategy using a pre-trained model optimized through self-play

These strategies represent a modern machine learning approach to the Prisoner's Dilemma, capturing complex patterns in game history through attention mechanisms rather than using hand-crafted rules.

The model processes the last 200 moves of both players, encoding game states (CC, CD, DC, DD) and using self-attention layers to identify patterns and make decisions. The implementation includes a complete neural network architecture with embeddings, attention layers, and classification components.

The pre-trained weights for the EvolvedAttention model are loaded from external data files.