Remove global ROOT_NODES/EDGES (merge after #86)

[qinsoon]

This PR is based on #86.

• Move stack scanning code to Rust, and implement Scanning::scan_roots_in_mutator_thread.
  • Implement necessary types/functions in Rust.
  • It is not necessary to implement this in Rust. However, it needs to use scan_gcstack(), which is implemented in Rust, and used for object scanning. It would be better to remove our implementation for it in C.
• Create scan_vm_specific_roots, and call it in Scanning::scan_vm_specific_roots.
  • Allow using RootsWorkFactory in C
• Remove global ROOT_NODES/EDGES as they are no longer used.
• Add assertions to check type size for our Rust and C types.

[udesou]

It looks good to me. This might require a few changes to support moving immix (tpinned roots), but I believe these changes introduced here aren't gonna impact on supporting tpinned roots.

[udesou]

Does this impact performance in any way? Have you tested it with the previous benchmarks? I'm just wondering whether it would help at least a bit since it removes those locks.

[qinsoon]

Does this impact performance in any way? Have you tested it with the previous benchmarks? I'm just wondering whether it would help at least a bit since it removes those locks.

I haven't tested that. It should improve the stack scanning. But I am not sure how much it matters for the whole GC time. I see this PR as a cleanup rather than an optimization. But anyway, I will run the benchmarks.

[qinsoon]

This is the result (compared with Julia GC). It seems we only see a clear improvement in pidigits. I think the reason is that stack scanning is a fixed cost per GC and we trigger much more GCs in pidigits than other benchmarks. So the improvement is only visible for pidigits.