When commits are backed out, we don't need to schedule the full set of tasks on the resulting push. We should only need to schedule the tasks that failed in the previous push(es).

This is difficult: If a test fails because of an unexpected behavior of the build, the build will also have to be requested.

if the test fails, won't we need to build+test jobs?

I think it depends on what is being backed out. If it's test-only changes, then we don't need the builds. If we're backing out source changes, then yeah, we need the builds too.