Implement $id and $ref properly

The current situation is a mess, and probably quite buggy. We have two separate schema crawlers - one in the canonizer, the other one in the type-building parser. We mix canonical refs and virtual refs too much.

Here's what I think a proper implementation should do:

* Have one type for traversing schema objects. It has to combine the schema object with its location, which can be any number of refs, I think. See the [relevant section in the JSON Schema spec](http://json-schema.org/latest/json-schema-core.html#rfc.section.8.2.4). The only way to traverse a schema is to use that type.

* Whenever a new schema file is loaded every single one of its refs is registered by the canonizer, so that no traversing is necessary to look up a `$ref`.

* When a `$ref` is resolved, the canonizer's map is consulted first. If the ref isn't registered, candidates for its address have to be attempted to fetch. There might be more than one, since a ref to `foo.schema` could mean `foo.schema` in the current directory, or relative to the where the referencing schema is. Alternative opinion: it can't mean the current directory, because then it would be dependent on where quicktype is called from. So either we've registered `foo.schema` before, via some `$id`, or we interpret it relative to the referencing schema.

The whole point of this is to decouple everything that's going on, and to have only one, clear place where policy is decided.