loader(safetensors): HfConfigExtractor + TransformerWeightsSafetensors for dense LLMs #166

Summary

Builds on #154 (safetensors binary parser foundation). Adds the HuggingFace config extractor + dense-LLM safetensors loader, enabling load of TinyLlama, Mistral, Phi-3, Qwen 2.5 family checkpoints.

Depends on

• loader(safetensors): binary parser foundation — SafetensorsFile / DType / TensorDescriptor #154 (binary parser foundation) — must merge first; this PR is stacked on it.

Proposed scope

• src/DotLLM.Models/SafeTensors/HfConfigExtractor.cs — parse HF config.json into ModelConfig (hidden size, layer count, RoPE config, vocab size, …).
• src/DotLLM.Models/Architectures/TransformerWeightsSafetensors.cs — load weights from a SafetensorsFile into TransformerWeights (GQA-aware, fused-QKV-aware).
• src/DotLLM.Models/ModelLoader.cs — LoadFromSafetensors(path, ModelConfig) entry point.
• TinyLlama-1.1B integration test (gated on fixture availability).

What lands in follow-up PRs (not this one)

• Multi-shard model.safetensors.index.json support (one PR).
• Per-architecture loaders: Phi-3 fused-qkv/gate_up split, Mixtral MoE, Qwen-MoE, DeepSeek-V2/V3.
• SmolLM3 / Gemma3 / Granite extensions.

Acceptance criteria

• HfConfigExtractor correctly populates ModelConfig for Llama / Mistral / Phi / Qwen architectures.
• TransformerWeightsSafetensors loads weights for a dense GQA Llama-class model end-to-end.
• ModelLoader.LoadFromSafetensors returns a usable model that generates a first token.
• Integration test exercises the load + first-token generation path.