Add Llama-3 / Mistral architecture support

[BioInfo]

Feature: Llama-3 and Mistral model support

Currently Dendrite supports LLaMA (2/1) and Qwen3 architectures. Llama-3 and Mistral are among the most commonly used open models.

What Changed in Llama-3

• GQA (Grouped Query Attention) — different num_kv_heads vs num_heads
• RoPE scaling: uses rope_scaling config field
• SentencePiece → tiktoken tokenizer

Mistral Differences

• Sliding window attention (SWA) in some layers
• Different FFN interleaving
• Same RoPE as Llama-2

Scope

1. model/llama3.rs — Llama-3 architecture (extend existing model/transformer.rs)
2. model/mistral.rs — Mistral architecture
3. Tests: golden output test against reference implementation (HuggingFace transformers)
4. Example: examples/llama3_inference.rs

Why This Matters

Most users asking "can Dendrite run model X?" will ask about Llama-3 first. This is a high-visibility gap.

Complexity

Medium — architecture is well-documented. GQA support may require changes to KvCacheConfig (num_kv_heads != num_attention_heads). Good entry point if you know transformer architectures.