MB5 PD ablation v2 tooling: conc completion-panel plot + gpu_monitor dep

- plot_pd_crossover.py fig_conc: lead with request-completion % (the honest
  collapse signal; latency percentiles count successes only), then mean-E2E /
  TPS; note PD-capped/colo-uncapped in the title.
- add microbench/fresh_setup/gpu_monitor.sh (referenced by the committed
  mb5_run_gpu.sh:73 for per-GPU util collection).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

microbench/fresh_setup/gpu_monitor.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+# Sample GPU utilization every 5s, output CSV
+# Usage: bash gpu_monitor.sh <output_file> [interval_s]
+#   Runs until killed (Ctrl+C or kill)
+
+OUT="${1:-/tmp/gpu_util.csv}"
+INTERVAL="${2:-5}"
+
+echo "timestamp,gpu,util_pct,mem_used_mb,mem_total_mb,power_w" > "$OUT"
+
+while true; do
+    TS=$(date +%s.%N)
+    nvidia-smi --query-gpu=index,utilization.gpu,memory.used,memory.total,power.draw \
+        --format=csv,noheader,nounits 2>/dev/null | while IFS=', ' read -r idx util mem_used mem_total power; do
+        echo "$TS,$idx,$util,$mem_used,$mem_total,$power"
+    done >> "$OUT"
+    sleep "$INTERVAL"
+done