Batched Trajectory Suffix Sum

Autopilot AI Foundation coding task over trajectory tensors: given input shaped `[batch, num_waypoint, 2]`, compute a suffix-style aggregate with output shaped `[batch, num_waypoint]`. Requirements Input tensor shape: [batch, num_waypoint, 2]. Output tensor shape: [batch, num_waypoint]. Compute a suffix-sum-style value over waypoints for each batch item. The exact target quantity was not fully leaked; prepare for cumulative distance-to-end, remaining displacement, or remaining scalar score variants. Notes The likely implementation pattern is vectorized reverse cumulative aggregation over the waypoint dimension. Clarify whether the aggregate is over raw coordinates, per-step displacement magnitudes, or an externally supplied per-waypoint scalar. For remaining path length, compute deltas between adjacent waypoints, convert each delta to a norm, then reverse-cumsum those segment lengths with a zero at the final waypoint. With NumPy / PyTorch, avoid Python loops over batch. Reverse the waypoint axis, apply cumulative sum, then reverse back. Preparation Implement three variants over a [B, T, 2] tensor: remaining displacement vector, remaining scalar path length, and suffix sum of a supplied [B, T] score tensor. Write shape tests for T=1, repeated waypoints, multiple batches, and float precision on long trajectories. Practice explaining why np.flip(np.cumsum(np.flip(x, axis=1), axis=1), axis=1) works and where segment arrays need padding.