I want to be able to transfer multiple files, and I want the sender to
control the file name so the receiver does not have to. This adds a step
in that direction by transferring the plan, but it breaks the actual
transfer. I'll fix that up next.

For a long time I was puzzled by an "illegal seek" error. It turns out,
the file was being closed becuase I dropped it immediately. The raw fd
is like a borrow, I need an owned fd.

We still need to open a final connection to break the sender out of the
accept() loop, but I'll do that in a follow-up commit.

It's a hack maybe, but I really want to avoid the async ecosystem.

Not sure if this is a good idea in the end or not, it harms debugging
but creates cleaner files, which on SSDs probably does not matter but
on spinning disks it can. But we use SSDs, so I am totally guilty of
optimizing prematurely here.

And that we target Linux only.