fix: run tools concurrently in PARALLEL ToolExecutionMode

[copybara-service]

fix: run tools concurrently in PARALLEL ToolExecutionMode

[kiwigitops]

The production change makes sense, but the new concurrency tests are still mostly wall-clock based. durationMillis < toolCount * sleepMillis proves the calls are not fully sequential, but it would also pass if a 5-tool run only executed two tools at a time, and it can get noisy under CI load.

A more direct regression would have SleepingTool record active-call count or wait on a shared latch/barrier, then assert that at least two tool invocations overlap before any of them are released. That would test the scheduler behavior without depending as much on timing margins.