Extend Context with RoPE Scaling

A model trained with RoPE on 4K tokens breaks at 16K, because it has never seen such large rotation angles. Two popular fixes:

• Position interpolation (linear): divide positions by a scale factor s, so 16K positions map into the trained 0–4K range.
• NTK-aware scaling: leave positions alone and raise the RoPE base to $b' = b \cdot s^{d/(d-2)}$. The highest frequency stays the same (local detail is kept) while the lowest frequency slows down by exactly s.