Chunking should always be primary-led

[eddyashton]

We currently have per-node chunking decisions, based on chunk_size options on the host configuration. This is dangerous - if it mismatches on different nodes, we'll get different chunks, breaking simple operator backups.

We also have a primary-led chunking path, where governance can trigger chunks, resulting in a flag at the start of a transaction that causes every node to produce a chunk at the same transaction.