Function bfs_parallel 

pub fn bfs_parallel(
    adj: &CsrMatrix,
    sources: &[NodeId],
) -> Result<Vec<Distance>>

Parallel BFS implementation (frontier-based).

Uses level-synchronous approach suitable for GPU parallelization:

1. Process all nodes in current frontier in parallel
2. Collect next frontier
3. Repeat until no more nodes to explore

On CPU, this uses rayon for parallelism (when available). The same algorithm structure maps directly to GPU kernels.