task graph: improve TestRunGraph tests

The test was a layered amalgam with multiple types of inputs (task
graph, paralellism, `errorOn` which ended up being used for arbirary
callback use cases) and expected outputs (most used by just one or two
testcases).

- document test inputs
- improve invariants to be better readable for individual testcases
- describe desired properties as invariants
- restructure `errorOn` as `callbacks` to be called for individual
  tasks rather than a single big function with `if` matching tasks

Author: petr-muller