== Doing

* Implement `columnar_2::save::save_document`



== Cleanups

* Add a `*Range` for each column type (including composite)
* For each of the `*Range` types add an `encode` method. Replace the ad hoc
  `*Decoder::encode` methods with these methods
* Get rid of `PrimVal`, do everything with `ScalarValue`. i.e. move the unknown
  type codes into `ScalarValue`.
* Make `Change` and `Document` directly hold a `*Columns` instead of
  `ColumnMetadata`, construct these at parse time
* Make `Chunk` an actual enum and parse the contents into that enum
* Change `ChangeOpColumns::encode` to `ChangeOpColumns::splice`, include a
  `replace_with` range, and create a new `encode` implementation which just
  calls `ChangeOpColumns::empty().splice(..)`
* `try_splice` on all encoders
* Remove rowblock, just have structs for `ChangeOps`, `DocChangeOps`,
  `DocChanges` which hold their own bytes. Maybe later we add a `GenericRows`
  thing but we seem to not really need it 

* Replace every use of `usize` as a lookup with a newtype wrapper