Set RUSTUP_CLOSE_THREADS=disabled to force single threaded IO, or
to a specific number if desired for testing/tuning.

This may improve rust-lang#1867, but has no impact on rust-lang#1866 due to the coarse
lock around the fd-handle table inside WSL.

Generalises the threaded IO closing to be fully threaded disk IO.

Adds a new environment variable to control tracing. Tracing is
manual and can be expanded. We can probably get better systemic
tracing from a suite of platform specific tools over time, but for
now this was low hanging fruit that allows pretty good insight
without too much code intrusion. Runtime overhead when disabled
is extremely low (I can unpack rust-docs on Linux in 1s).

This was only needed because we were not controlling the permissions
accurately during unpack, which can now be corrected.

When directories complete, start writing the files/dirs within
that directory that have been decompressed already.

Avoids a stat() + create_dir_all() in the main thread permitting
more concurrent IO dispatch in exchange for memory pressure.

Heavily deferred directories can lead to very large amounts of RAM
holding pending IOs, and under low memory limits this caused failures.

Reduce the stack size for IO worker threads to 1MB, as the default
of 8MB on Linux is wasteful for IO worker threads.