Broker load is not cancelled even if some instances already failed

**Describe the bug**
When processing a long run broker load, and some BE crashed, that broker load job may wait until job timeout to cancel.

This will also happen in `Insert` and `Export` process.

**How to Fix**
Use a special CountDownLatch to directly count down to zero if some fragment instances reporting failure. So that all observers waiting for this CountDownLatch will return.
