local training - Jordan's model

[SteveAndCow]

This pull request introduces significant improvements to the local training pipeline for chess models, including a new training script, enhancements to the dataset processing, and updates to the transformer model architecture. The main themes are the addition of a training script, improvements to dataset handling, and architectural changes to the transformer model.

Training Pipeline Enhancements:

• Added a new script local_training.py that sets up and runs a training loop for chess models using PyTorch, including model saving, evaluation, and dataloader setup. This script enables local training using either the ChessCNN or ChessTransformer models.

Dataset Handling Improvements:

• Refactored fen_to_bitboards in stockfishdataset.py to output a 12x8x8 tensor for each board position, supporting both white and black to move, and added a StockfishDataset class compatible with PyTorch’s Dataset API. This makes data loading and preprocessing more robust and efficient for model training.
• Updated the dataset script to process input files more efficiently, handle mate scores, normalize evaluation values based on the side to move, and save processed tensors for both bitboards and evaluations.

Transformer Model Architecture Updates:

• Introduced a custom MultiheadAttention module that uses Flash Attention for improved efficiency, although it is not yet integrated into the encoder.
• Increased the number of layers in the ChessTransformer model from 4 to 6, potentially improving model capacity and performance.
• Commented out the use of relative positional bias and its associated attention mask in the transformer encoder layer, simplifying the attention mechanism for now.