Do fix the reftable problems when writing refs massively concurrently

[dscho]

I spent essentially an entire work day trying to figure out how to fix the t0610.47 failure in Git v2.47.0-rc1 last Friday, hoping to get it fixed before releasing the corresponding Git for Windows v2.47.0-rc1. But at some stage I had to stop because it became apparent that that was not to be.

However, this issue should not be forgotten. Here is the latest workflow run where I attempted not only to fix the issue but also to investigate why using POSIX semantics was not enough (my hunch is that the problem was no longer mingw_rename() and I was simply barking up the wrong tree).

[dscho]

For the record, this is still not fully fixed, as the CI builds still occasionally fail despite the work-around that @pks introduced.