v0.40.0

Fixed

• APR→GGUF export no longer produces corrupt GGUF for AprQ8 tensors: export
  silently mapped APR-native AprQ8 (single-whole-tensor-scale 8-bit,
  [f32 scale] + [i8×N] = 4+N bytes) to GGML Q8_0 (per-32-block,
  ceil(N/32)·34 bytes) and emitted the raw APR bytes unconverted under the
  Q8_0 label — a corrupt GGUF that any llama.cpp loader misreads (reachable via
  apr import x.gguf && apr export --format gguf on Q4_K_M models). Export now
  rejects AprQ8 with a clear error (pointing to apr convert → F32/F16),
  restoring import/export symmetry (the import side already refuses GGUF Q8_0,
  and AprQ4 export was already rejected). Layout-identical dtypes
  (F32/F16/Q4K/Q6K) export unchanged. Contract
  contracts/apr-gguf-export-symmetry-v1.yaml (FT-APRQ8-001/002).