Decouple root scanning from ProcessEdgesWork.

[wks]

This draft PR is a proposal for changing the root-scanning interface.

When scanning roots, the VM should only report what roots it has, not deciding what work packets to create. But currently, scan_thread_root, scan_thread_roots and scan_vm_specific_root take a generic type parameter <E: ProcessEdgesWork> which binds them to a ProcessEdgesWork. This may not be general enough, in two ways:

1. Some VMs cannot present roots as edges. They can only present roots as nodes because they cannot update the root edges. One example is the Ruby VM. It uses conservative stack scanning, and cannot update roots. It gives the GC a list of values on the stack, and some of them are roots (can be filtered using the is_mmtk_object function in the core).
2. Some GC algorithms do root scanning significantly differently than our current ProcessEdgesWork. Although ProcessEdgesWork is still a trait, it is too tied to some kinds of GC. For example, it contains the logic of tracing objects and scanning objects (via ScanObjects). The real culprit is the ProcessEdgesWork.

However, from the code in this PR, we observe that we need to wrap ProcessEdgesWork in such a complicated manner just to get a chance to call the plan-specific trace_object function. Probably it is easier to do this refactoring if we eliminate (or weaken) the ProcessEdgesWork trait first.

[wks]

Superseded by #611