Investigate replacing O(n) locking with refcounting tasks in taskgroup killing

[bblum]

Described in a comment on #3098. Worth making a perf test about. Requires adding two runtime calls for ref-ing and deref-ing task pointers.

[eholk]

You might be able to avoid the runtime calls by using the atomic intrinsics that I added to Rust.

[bblum]

I'd rather not... *rust_task is an opaque pointer type for good reason (and there's already an atomic refcount inside).

[bblum]

Conclusion: Not worth it. Significant increase to design complexity (easier to introduce a memory-management bug), and performance gain not outweighed by orthogonal performance hit.

Performance with task-perf-jargon-metal-smoke, where the perf win doesn't show up, exhibiting the perf hit: http://bblum.net/images/perf-refcount-vs-exclusive/100-generations.svg
Performance with task-perf-linked-failure, where the perf win shows up and seems to balance the hit: http://bblum.net/images/perf-refcount-vs-exclusive/linked-failure.svg