tests/ep: hoist combine output tensor out of the timed loop

The LL combine benchmark was cloning the ~58 MB dispatch recv buffer
('recv_x.clone()') on every timed iteration, adding ~20 us of D2D
memcpy per sample and masking kernel-level changes. It also called
torch.empty() for the output inside the loop. Both now live outside
the timed region; the kernel is invoked against a persistent bench_out
and the recv_x produced by the most recent dispatch.

test/python/ext/ep/test_low_latency_multirank.py

@@ -221,11 +221,16 @@ def main():
             False, False, False,  # use_fp8, async, return_recv_hook
         )
 
-    def _combine(dout):
+    # Hoist combine's output-tensor allocation out of the timed loop so the
+    # measurement reflects the kernel cost. (The original test also cloned the
+    # ~58 MB dispatch recv buffer on every iter, adding ~20 us of D2D memcpy
+    # to each combine sample and masking kernel-level changes.)
+    bench_out = torch.empty((num_tokens, hidden), dtype=torch.bfloat16, device="cuda")
+
+    def _combine(dout, out_):
         (recv_x, _scales, _cnt, src_info_, layout_range_, _ev, _hk) = dout
-        out_ = torch.empty((num_tokens, hidden), dtype=torch.bfloat16, device="cuda")
         buf.low_latency_combine(
-            recv_x.clone(), topk_idx, topk_weights,
+            recv_x, topk_idx, topk_weights,
             src_info_, layout_range_,
             num_tokens, num_experts,
             False, False, False,
@@ -233,7 +238,7 @@ def main():
         )
 
     for _ in range(warmup):
-        _combine(_dispatch())
+        _combine(_dispatch(), bench_out)
     torch.cuda.synchronize()
     dist.barrier(group=group)
 
@@ -251,7 +256,7 @@ def main():
     dist.barrier(group=group)
     start_ev.record()
     for _ in range(iters):
-        _combine(dout)
+        _combine(dout, bench_out)
     end_ev.record()
     torch.cuda.synchronize()
     comb_us = start_ev.elapsed_time(end_ev) * 1e3 / iters