Process creation on Windows is needlessly serialised

When `close_fds = True`, we can explicitly pass the processes that the child is allowed to inherit. This completely avoids the problem of [GHC bug #2650](https://gitlab.haskell.org/ghc/ghc/-/work_items/2650) that made locking necessary.

In a microbenchmark, spawning 200 processes with `close_fds = True` concurrently takes about 1.5s on my machine, with over 40% of the time (0.7s) wasted due to waiting on the lock.

While working on speeding up GHC's downsweep pass, I found that on Windows a lot of time was spent on process spawning (e.g. to run the C preprocessor), and that this serialisation was a significant component of that. In a rebuild of the `vulkan` package, the downsweep pass spawns a huge amount of subprocesses, and removing this locking takes the time spent in downsweep from 31s to 27s (4s total timesave).

Start by tracing the Windows process-spawning path for close_fds=True and review the explicit child-inheritance approach described in the issue, alongside GHC bug #2650. Validate the change with the concurrent process-spawning microbenchmark described here; done means the unnecessary lock serialization is removed without changing which handles the child inherits.

Written by the indexing model from the issue text.