Support nested schemas

[jcrobak]

Rather than flattening schemas and adding a '.' between paths
in the schema (e.g. foo.bar), support the schema path as a
first-class object (for schema operations, at least). This
is an experimental implementation and likely has bugs. But it
supports some simple cases.

This implementation changes behavior. Specifically:

• DictReader() now has a flatten argument that defaults to
  False. If Flatten is false, DictReader will read nested data
  as {'foo': {'bar': 1}} instead of as {'foo.bar': 1}.
• Likewise, this is the new default behavior for the command-line
  tool with --format json. This can be changed with --flatten.

Known issues:

• Repetition-levels still aren't supported. A file with arrays
  will break.
• nulls aren't interpretted at the level. (e.g.: {"foo": null}
  will be interpetted as {"foo": {"bar": null}} if foo has
  a child of bar.

Refs: #32

[martindurant]

"Read repetition levels before definition levels." - thank you for pointing this out!