[FEATURE] Modernize BGL shortest path algorithms

[Becheler]

The Sage Math graph module uses Boost Graph. This quick exploration comes from kind feedback from its maintainer David Coudert:

The BGL currently implements traditional (~1990) algorithms for shortest paths:

• dijkstra_shortest_paths : 1959 single source, single shortest path tree, but constrained to positive weights
• bellman_ford_shortest_paths : 1958 extends to negative weights with negative cycle detection
• johnson_all_pairs_shortest_paths : 1977 combines bellman for reweighting with dijkstra running for each vertex
• floyd_warshall_all_pairs_shortest_paths : 1962 better than johnson for dense graphs, uses dynamic programming
• astar_search: 1968 single source to single target path, heuristically

But it lacks more modern algorithms:

• Yen's algorithm (Yen 1971, cited ~500x) : k-shortest simple (loopless) paths between two vertices. At each iteration, finds the next shortest path by deviating from previously found paths. Baseline must-have
• k-Shortest Simple Paths, Postponed Node Classification algorithm variant, (Al Zoobi et al 2023, cited ~13x) : best speed/memory tradeoff for finding k alternative paths, most performant
• k-Shortest Simple Paths, SB* variant, (Al Zoobi et al 2023, cited ~13x) : fastest for small k, higher memory
• Contraction Hierarchies (Geisberger et al., 2008, cited ~1200x) : amazing for road networks, 1000x+ speedup over Dijkstra
• Hub labelling ( Abraham et al., 2011, cited ~400x) : sub-milliseconds shorttest path length oracle: avoids running dijsktra millions of times

[jeremy-murphy]

At the moment, I think merging PBGL into BGL is a higher priority, as using modern parallelization would presumably be a huge improvement to the scaling of the old algorithms.