Perhaps built-in Schema implementations can re-use buffers internally to reduce memory allocations for each `Decode` call?