The CLI resolver always treats type keys with the name of `main` as
top-level objects. For example, `router:main` resolves to `app/router.js`
instead of `app/routers/main.js`.

This change updates the default resolver to always look for
`foo-bar:main` as `App.FooBar`. This is in anticipation of application
init/boot cleanup that we are doing as part of the FastBoot effort.

Instead of storing a `__container__` on the Application, this commit
stores an ApplicationInstance, which manages the per-instance lifecycle.

The next step will be allowing multiple `ApplicationInstance`s to exist
at once, enabling a single app to serve multiple FastBoot requests at a
time.

This is a conceptual improvement that brings application “reset” in line
with the destruction infrastructure.

As part of this commit, we eliminated an ad-hoc call to router.reset(),
allowing that logic to happen as a natural consequence of destruction
of the `ApplicationInstance`.

In general, the goal of this commit is to move all responsibility for
application state into an object that manages it.

Based on feedback from @stefanpenner, we moved responsibility for
creating the default container from the `Application` to the
`ApplicationInstance`. Instead, the instance points back at the
application’s registry, using it as the basis for a new container.

This commit also adds initial documentation to the application instance.

This commit introduces a new (feature flagged) API for adding
instance initializers.

Instance initializers differ from normal initializers in that they are
passed the app instance rather than a registry, and therefore can access
instances from the container in a safe way.

This design not only allows us to avoid expensive app setup for each
FastBoot request, it also minimizes the amount of work required
between acceptance test runs (once the testing infrastructure
is updated to take advantage of it).

This commit also removes a previously introduced deprecation that was
not behind a feature flag. That deprecation (when emitted with the
feature flag enabled) now points to a comprehensive deprecation guide.

Thanks @stefanpenner!