Add sum_scatter to the ring backend

[erwinzhang7]

Fixes part of #4285.

all_reduce_impl already runs its send/recv plan twice, and the code says why:
"Two times the same send/recv operations, first scatter reduce and then gather."
The k == 0 half is a reduce-scatter, so sum_scatter is that phase with the local
segment kept instead of continuing into the gather.

Verified on two M4 Pro minis over a direct Thunderbolt link: the call returns
cleanly on both ranks where current main hangs indefinitely, and the ring
distributed test suite gains a sum_scatter case.

Note that this does not address the second half of the issue. An error raised out of
ReduceScatter::eval_cpu still does not reach the caller; this change means the
common path no longer raises at all.