🪟 Add sliding window iteration mode for datasets

[dariocazzani]

What does this PR do?

This PR introduces a sliding window iteration mode for dataset processing, providing an alternative to the existing chunking approach. Sliding windows create overlapping sequences that move by one token at a time, significantly increasing the number of training samples and improving context coverage.

Details

• Add iteration_type parameter to training config with "chunking" or "sliding" options
• Implement SlidingWindowDataset class for creating overlapping token windows
• Update HFDataSource to handle both iteration modes via pattern matching
• Add comprehensive test suite covering sliding window functionality
• Sliding mode disabled for streaming datasets due to sequential constraints
• Bump version to 0.5.0 to reflect new feature addition

Highlights

The key difference between iteration modes:

# Chunking: Non-overlapping blocks
# Input: "0123456789" with block_size=3
# Samples: [012], [345], [678]

# Sliding: Overlapping windows (stride=1)  
# Input: "0123456789" with block_size=3
# Samples: [012], [123], [234], [345], [456], [567], [678]

Usage in config:

training = ScratchGPTTraining(
    iteration_type="sliding",  # or "chunking" (default)
    # ... other params
)

[ayeganov]

LGTM!