Remove cancellation workflow, fail-fast. (#1415)
A flaky test can bring down an entire set of CI runs but I only notice when I go back to look at the results, which means I've already waited for what I instinctively feel is the entire duration of the tests. Cancelled runs give me *no* information about the change.
Like this run here:
https://github.com/vercel/turborepo/runs/6934131393?check_suite_focus=true
I only wanted to make sure it passed Windows CI, but that got cancelled for reason of a flaky test on Linux.
This removes our cancellation workflow (replaced by built-in `concurrency` groups), and disables fast fail.
It also corrects a bug in caching.
