We need the logging handler setup before creating the Stream(),
as the Stream() will expect logging to be already functional.

It makes not sense to type `bst fetch --track --deps all <targets>`,
because tracking will inevitably modify the build plan.

Stream initialization will not cope with this either, instead of
silently doing something which does not make any sense, we add an
assertion that this should not happen.

Unfortunately since `plan` is the default deps type for `bst fetch`,
this is likely to happen so it's important to correct.

This patch adds a warning in the case tracking of the build plan
elements is requested, and converts the request to track all elements
instead.

Here the pipeline becomes essentially stateless, some dangling
state remains to be factored out because of frontend accesses
which will be changed in a later commit.

Essentially, the Pipeline.load() method no longer has any knowledge
of the specific purposes of the loaded targets, and now takes
a list of target groups and returns a corresponding list of element
groups.

The Stream() business logic methods now use other pipeline helper
methods to create and filter lists from the loaded target elements.

The Stream() also finally absorbs the Scheduler frontend facing
APIs. However Queues are still exposed on the Stream object for
logging purposes and through callbacks such that the frontend can
retry elements.

This removes some additional initialization code from Pipeline().

Some symbols have changed here, the initialization is now called
from Stream(), and a test case was also adjusted for this.

There is no more need for this distinction.

This was previously decided in CLI, but knowledge of what to initialize
has been moved to Stream().

Now there is no more point to store this configuration in the Context,
we just have the Stream() decide it when asking the Pipeline() to
invoke the Loader().