Second model family with perplexity check (Llama 3.1 8B, Mistral 7B)

[parda21]

Everything so far is Qwen2.5 / Qwen3. Convert Llama 3.1 8B and Mistral 7B, run scripts/bench_model.py with perplexity on WikiText-2 against native, report size, tokens/s and that the perplexity is identical.