Use a real (probing-generated) scorer in benchmarks

Until now, our routing benchmarks used a synthetic scorer,
generated by scoring random paths to build up some history. This is
pretty far removed from real-world routing conditions, as
alternative paths generally have no scoring information and even
the paths we do take have only one or two past scoring results.

Instead, we fetch a static serialized scorer, generated using
minutely probes. This means future changes to the scorer's data may
be harder to benchmark, but makes for substantially more realistic
benchmarks for changes which don't impact the serialized state.