Support more model architectures

`pooji_vllm/model.py` implements the Llama-style decoder stack and rejects anything else with a clear error. Tested on Qwen2.5-0.5B, SmolLM2-135M, TinyLlama-1.1B. **Not yet supported:** Gemma (different normalization and embedding scaling), Phi-3 (fused QKV projection), MoE models (expert routing in the MLP), Mamba-style layers (no KV cache at all). **Where:** `Model.__init__` for config parsing, `Model.qkv` and `Model.mlp` for the per-layer math. **Good first issue** if you pick a single architecture and add it with a test in `tests/test_model.py`.