[ML] Read one object at a time for incremental training state

Currently, we read one big document for the data summarisation and one big document for the inference model, but there is nothing to stop us reading individual state documents one at a time using the `doc.ParseStream` API. This would significantly cutdown the memory when of restoring state for incremental train since we only inflate the state on-the-fly when reading the dechunked stream. Relates to #1849.