Work with double/floats as rounded integers

[knutwalker]

Instead if using plain doubles, we can multiply every value by a constant factor—e.g. 1_000_000—and implement the algorithms on those integers and divide back to doubles when writing the results back.

• ints use only 4 bytes instead of the 8 bytes for doubles
• we lose precision and range of representable value (since we only have half the data available)
• arithmetic on doubles is more expensive compared to arithmetic on integers

There should be JMH benchmarks to show the performance gains, if any.

[mknblch]

We would benefit from having integer weights in: org.neo4j.graphalgo.impl.ShortestPathDeltaStepping