Add production multi-GPU training scripts (B=2 + compile + shards)

scripts/launch_4gpu.sh

@@ -0,0 +1,12 @@
+#!/bin/bash
+# Launch Fractus 1B training on 4 GPUs
+set -e
+cd "$(dirname "$0")/.."
+mkdir -p logs checkpoints
+
+for i in 0 1 2 3; do
+  echo "Starting GPU $i..."
+  GPU_ID=$i CUDA_VISIBLE_DEVICES=$i setsid python -u scripts/train_1b_multi_gpu.py > logs/gpu$i.log 2>&1 < /dev/null &
+done
+
+echo "All 4 GPUs launched. Monitor with: tail -f logs/gpu*.log"