Applications of traits with `@idRef` members don't create relationships
between shapes, so removing unreferenced shapes from the following:
```
service FooService {
    ...
    operations: [GetFoo]
}

operation GetFoo {
    output := {
        foo: Foo
    }
}

@trait
@idref(failWhenMissing: true, selector: "*")
string myIdRef

@myIdRef(Referenced)
structure Foo {}

stucture Referenced {}
```
will remove `Referenced` from the model. This commit changes
UnreferencedShapes so that `Referenced` is considered connected
to `FooService` because it is referenced by an `@idRef` on a shape
connected to the service.

Determining which shapes are connected via id ref is done by the
new IdRefShapeReferences class, which uses the following process:
1. Get all trait definition shapes.
2. Find all paths from those trait definition shapes to shapes with
the idRef trait. Each path found says that applications of the trait
definition may have a shape reference.
3. For each path, get all applications of the trait on connected shapes
and search the node value using the path to the shape that had idRef.
4. For each search result, exclude all connected shapes from being
unreferenced.

Other changes:
- A few test cases that expected to receive an UnreferencedShape
event were updated
- An `asShapeId` method was added to StringNode

Previous impl wouldn't consider shapes referenced through multiple
levels of idRef as connected, since idRef'd shapes could themselves
be connected to other shapes that have idRef references. Refactoring
to a custom NeighborProvider means the shape walker just traverses
those relationships in a single pass.

Also makes most NodeQuery queries static instances, and
moves some stuff around in tests.

Changes NodeQuery to use a single instance of a Queue for keeping
track of and aggregating query results to avoid extra allocations
within queries and when joining the results.