fix(agentic-mtp): stop dividing CUDA_GRAPH_MAX_BS by TP for DPA

Remove the per-DP-rank division of CUDA_GRAPH_MAX_BS — keep it at CONC
so decode batches up to the concurrency level stay on the CUDA graph
fast path (was dropping from 52 to 7 with TP8).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: seungrokj

benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh

@@ -113,10 +113,8 @@ fi
 # internally. CUDA graph capture is per scheduler, so only its batch size is
 # divided across DP ranks.
 MAX_RUNNING_REQUESTS=$CONC
+# 0719
 CUDA_GRAPH_MAX_BS=$CONC
-if [ "$DP_ATTENTION" = "true" ]; then
-    CUDA_GRAPH_MAX_BS=$(( (CUDA_GRAPH_MAX_BS + TP - 1) / TP ))
-fi
 [ "$CUDA_GRAPH_MAX_BS" -gt 128 ] && CUDA_GRAPH_MAX_BS=128
 
 # Simulated acceptance-length (AL) settings.