feat: huoyi M20 smooth/bridge training and stability docs

Current code snapshot: smooth/bridge task configs, smooth random terrain,
roll/pitch stability rewards, M20 URDF updates, eval profile and stability
training documentation.

scripts/reinforcement_learning/rsl_rl/eval.sh

@@ -0,0 +1,23 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+TASK=${TASK:-Rough-Deeprobotics-M20-v0}
+CHECKPOINT=${1:-${CHECKPOINT:-}}
+NUM_ENVS=${NUM_ENVS:-250}
+DEVICE=${DEVICE:-cuda:0}
+EVAL_TYPE=${EVAL_TYPE:-all}
+
+if [[ -z "${CHECKPOINT}" ]]; then
+  echo "Usage: $0 <checkpoint.pt>"
+  echo "Or set CHECKPOINT=/path/to/model.pt"
+  exit 1
+fi
+
+echo "Running eval_type=${EVAL_TYPE}"
+python scripts/reinforcement_learning/rsl_rl/eval.py \
+  --task="${TASK}" \
+  --checkpoint="${CHECKPOINT}" \
+  --num_envs="${NUM_ENVS}" \
+  --device="${DEVICE}" \
+  --eval_type="${EVAL_TYPE}" \
+  --headless