This properly decouples data buffer nodes from data buffers and should
address some subtle bugs, especially with uninitializing data buffers
while they're still loading.

This also cleans up the decoding code, in particular by avoiding some
duplicate code between the synchronous and asynchronous decoding paths.