Fix RS graph capture test reference — use all_reduce not reduce_scatter_tensor

iris RS takes (M,N) input and (M,N) output (same shape). Each rank
reduces its assigned tiles from all peers. The torch equivalent is
all_reduce, not reduce_scatter_tensor (which has different shape
semantics: ws*M input → M output).

Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>

Author: mawad-amd

tests/ccl/test_graph_capture.py

@@ -96,8 +96,10 @@ def collective(out, inp):
             ctx.ccl.reduce_scatter(out, inp)
 
         def reference(out, inp):
-            big = inp.expand(ws, -1).contiguous().view(ws * M, N)
-            dist.reduce_scatter_tensor(out, big)
+            # iris RS: all ranks have same-shaped (M,N) input, each rank reduces
+            # its assigned tiles. Equivalent to all_reduce (sum all inputs).
+            out.copy_(inp)
+            dist.all_reduce(out)
 
         _graph_capture_test(ctx, collective, reference, (M, N), (M, N))