Introduce arc_agi_manager for loading and tokenizing ARC-AGI reasoning tasks.

This implementation provides:

• JSON-type parsing for ARC-AGI task files (training and evaluation sets)
• Grid tokenization with special markers for implicit dimension encoding
• Sliding window context generation for causal language model training
• Detokenization utilities for grid reconstruction from token sequences
• Serialization/deserialization support for faster dataset loading
• Task access by index or ID

The tokenization strategy encodes variable-size grids (1x1 to 30x30) using row-end markers, enabling transformer models to learn grid dimensions without explicit specification.
Context windows implement left-padding for proper causal attention masking.