v0.39.0

Added

• BF16 (bfloat16) GGUF loader support: BF16 GGUFs (ggml type 30) now load.
  Previously any BF16 GGUF hard-failed with "Unsupported quantization type: 30"
  because get_tensor_f32 (embeddings/norms/lm_head) and tensor_byte_size
  (per-layer weights) lacked a BF16 dispatch arm — even though the matmul weight
  path already consumed BF16. The fix adds the two arms (reusing the existing
  simd_bf16_to_f32 converter; BF16 is 2 bytes/elem, value from_bits(b << 16))
  and GGUFBuilder::add_bf16_tensor for fixtures. Contract
  contracts/bf16-dequant-v1.yaml (FT-BF16-001 golden converter + FT-BF16-002
  end-to-end dispatch). Dense BF16 load path complete (get_tensor_f32 +
  tensor_byte_size + matmul); MoE/CUDA BF16 remain follow-ups.