The OSS dispatcher already aborts the prior dispatcher build for the same
PR, but the 8 leaf jobs it spawned via `build job: ..., wait: true` keep
running until they finish naturally. That ties up build executors every
time someone pushes a new commit to a PR.

Port the pattern NIXL uses: scrape each stale dispatcher's console log
for `Starting building: <name> #<num>` lines to find its children, kill
the leaves first so the dispatcher's wait unblocks, then kill the
dispatcher. Use `doKill()` instead of `doStop()` for a hard stop, and
retry up to 3 times with a 5s gap so builds caught mid-startup still get
torn down.

Failures inside the abort block are still swallowed and logged - we
never want stale cleanup to break a fresh build.