Start by moving the JobCanceler here.

The point of making a sub-crate is to ensure that only the necessary
parts of a program get compiled; if a dependent package needed only
subsystem-util, and not subsystem, then subsystem wouldn't need to
be compiled.

However, that will never happen: subsystem-util depends on
subsystem::messages, so subsystem will always be compiled.

Therefore, it makes more sense to add it as a module in the existing
crate than as a new and distinct crate.

This struct can be constructed when the local node is a validator;
the constructor fails otherwise. It stores a bit of local data, and
provides some utility methods.

This means that the work of implementing a subsystem boils down
to implementing the job, and then writing an appropriate
type definition, i.e.

pub type CandidateBackingSubsystem<Spawner, Context> =
	util::JobManager<Spawner, Context, CandidateBackingJob>;

This turns out to be relatively complicated and requires some
unsafe code, so we'll want either detailed review, or to choose
to revert this commit.

This works within `util.rs`, but fails in `core/backing/src/lib.rs`,
because we don't actually create the struct soon enough. Continuing
down this path would imply substantial rewriting.

This reverts commit a5639e3.

The fact is, the new API is more complicated to no real benefit.

Now the only public types exposed in that module are
CandidateBackingSubsystem and ToJob. While ideally we could reduce
the public interface to only the former type, that doesn't work
because ToJob appears in the public interface of CandidateBackingSubsystem.

This also involves changing the definition of CandidateBackingSubsystem;
it is no longer a typedef, but a struct wrapping the job manager.