Updating to latest llama.cpp

[AshtonOhms]

Hi! I updated the bindings to be compatible with the latest head of llama.cpp master - enough to get the Main.hs example working again.

Summary:

• A few functions no longer exist in llama.h: Namely, the *FromModel functions
• The old sampleRepetitionPenalty and sampleFrequencyAndPresencePenalties functions are now replaced with the sole sampleRepetitionPenalties function
• New function in wrapper.c for new_context_with_model
• llama.h's token_to_str has been replaced with token_to_piece. A corresponding wrapper has been added to the Main.hs example

Tested by running the Mixtral MoE model, and it seems to be working properly.
stack run examples -- -m ../mixtral/mixtral-8x7b-instruct-v0.1.Q2_K.gguf -p "Hello, my name is " -t 12

I probably should do a version bump of some kind, but I'm not sure - let me know your thoughts. Cheers!