docs: multi-GPU Newton training guide

Author: ClawLabby

docs/source/how-to/multigpu_newton_training.md

@@ -0,0 +1,41 @@
+# Multi-GPU Training with Newton Physics
+
+## Overview
+
+This fix enables multi-GPU distributed training with Newton physics and camera
+rendering in Isaac Lab. Three interrelated bugs were fixed:
+
+1. **Device assignment**: Non-rank-0 GPUs were assigned to wrong CUDA devices
+2. **Camera renderer**: Newton warp renderer preset wasn't applied correctly
+3. **Fabric mode**: Non-cuda:0 GPUs need fabric disabled to avoid race conditions
+
+## Usage
+
+```bash
+NCCL_P2P_DISABLE=1 NCCL_IB_DISABLE=1 \
+torchrun --nproc_per_node=4 \
+  scripts/reinforcement_learning/rsl_rl/train.py \
+  --task Isaac-Dexsuite-Kuka-Allegro-Lift-v0 \
+  --num_envs 1024 \
+  --headless --distributed \
+  presets=newton,cube
+```
+
+## Environment Variables
+
+- `NCCL_P2P_DISABLE=1`: Required for containers with IOMMU (P2P transport fails)
+- `NCCL_IB_DISABLE=1`: Required when InfiniBand is not available
+
+## Known Issues
+
+- CCD solver warnings (`opt.ccd_iterations needs to be increased`) are expected
+  under high contact scenarios. The NaN watchdog handles resulting physics failures.
+- `abnormal_robot` termination rate is higher with camera observations (~50%)
+  vs state-only (~22%). This is a physics stability issue, not a training bug.
+
+## Testing
+
+```bash
+# Run multi-GPU Newton physics tests
+python -m pytest source/isaaclab/test/test_multigpu_newton.py -v
+```