Implement SuperHOT/interpolated RoPE support

[philpax]

Another llama.cpp feature that seems to have shrunk the paper-to-implementation pipeline to less than one week!

This allows for a much longer context (assuming you have the (V)RAM for it)

• Extending context size via RoPE scaling ggerganov/llama.cpp#1965
• Modified RoPE with linear scaling ggerganov/llama.cpp#2019
• Implement customizable RoPE ggerganov/llama.cpp#2054

We can probably close out #77 if this is done.

[LLukas22]

To do this we only need a new rope_scaling model parameter. Or am i missing something?