♟️ Add comprehensive chess training example with domain tokenizer

[dariocazzani]

What does this PR do?

This PR adds a complete chess move prediction example that demonstrates training transformers on domain-specific sequential data. It includes a Lichess database downloader, PGN parser, chess-specific tokenizer, and training pipeline that learns chess patterns from grandmaster games without knowing the rules.

Details

• Add examples/chess.py: Complete chess training pipeline with Lichess database downloading, PGN parsing, and move generation demo
• Add examples/chess_tokenizer.py: Domain-specific tokenizer for Standard Algebraic Notation with comprehensive chess vocabulary (~10k tokens)
• Add comprehensive test suite for ChessTokenizer with save/load, move encoding, castling, and promotion testing
• Update README with instructions for installing optional example dependencies
• Add zstandard as optional dependency for chess database decompression

Highlights

The chess tokenizer creates a deterministic vocabulary covering all valid chess moves:

@staticmethod
def _create_vocabulary() -> list[str]:
    """Generates the complete, deterministic vocabulary for chess."""
    tokens = {"[PAD]", "[UNK]", "[BOS]", "[EOS]", "*", "+", "#"}
    tokens.add("O-O")    # Kingside castling
    tokens.add("O-O-O")  # Queenside castling
    
    # Generate all possible piece moves, captures, and promotions
    for piece in ["N", "B", "R", "Q", "K"]:
        for square in squares:
            tokens.add(piece + square)  # e.g., Nf3
            tokens.add(piece + "x" + square)  # e.g., Nxd4

The training pipeline demonstrates domain adaptation with chess-optimized configuration:

def create_chess_config(tokenizer_vocab_size: int) -> ScratchGPTConfig:
    architecture = ScratchGPTArchitecture(
        block_size=256,        # Longer context for chess games
        embedding_size=384,    # Balanced for chess vocabulary
        num_heads=8,           # Good attention for chess patterns
        num_blocks=6,          # Sufficient depth for chess understanding
        vocab_size=tokenizer_vocab_size,
    )

[ayeganov]

Looks good, needs a few touch ups.