nx_f32_llm_read_dims.nx

nx_f32_llm_read_dims.nx -- read model dims from GGUF metadata. Walks the GGUF metadata section to derive model dimensions and writes them into NxF32LlamaModel. Closes the "caller pre-fills dims" gap from nx_f32_llm_load_weights_from_gguf. Architecture awareness: Reads "general.architecture" string; uses it as the key prefix for "<arch>.block_count" etc. Supports qwen2, llama (Llama-2), llama (Llama-3 same key namespace), and any other arch that follows the standard <arch>.{block_count,embedding_length,...} convention. Keys read: general.architecture (string) <arch>.block_count (u32 -> n_layers) <arch>.embedding_length (u32 -> hidden_dim) <arch>.attention.head_count (u32 -> n_heads) <arch>.attention.head_count_kv (u32 -> n_kv_heads, optional; defaults to n_heads if missing -- the no-GQA case) <arch>.feed_forward_length (u32 -> ffn_dim) vocab_size derived from token_embd.weight's dim_0. head_dim = hidden_dim / n_heads. genealogy_id: gguf_v3_metadata_spec_gerganov_2024 + llama_qwen_arch_keys lineage_id: substrate_f32_llm_read_dims_v1