Add automatic flake detection

[ds300]

A big problem with turborepo is that when tasks end up running too often, you get no feedback as to why.

Lazyrepo already helps out here by listing what changed from one run to the next.

Another potentially very useful thing to do would be to detect and flag common causes of flakiness.

The biggest cause is when a task modifies one of its own input files, or the input files of one of its upstream dependencies.

It would be fairly minimal overhead to check the mtimes of upstream input files after each task ends. I wonder whether fs.watch would be cheap enough for this purpose too.