Space Improvement

The $3D$ DP state dp[k][i][j] uses $O(n^3)$ space. You can reduce this to $O(n^2)$. Notice that dp[k+1][i][j] only depends on dp[k]. You can update dist[i][j] in place without storing all $k$ layers.

This is why Floyd-Warshall uses a $2D$ array and updates it during iteration. The space drops from $O(n^3)$ to $O(n^2)$. This improvement makes the algorithm practical for graphs with thousands of vertices.