I did consider avoiding locks, but as long as the close() in VCA_Shutdown()
happens concurrent to vca_acct(), the file descriptor could become invalid
at any time, so even if we'd not hit the sock > 0 assertion, other errors
could occur / assertions be hit.

We could also move the close to vca_acct() and just signal it to finish off,
but this would delay the close by the sleep(1), unless we further complicated
things with a condvar.

Avoiding the lock would require us to make any operation on a socket fd
ignore errors, which is totally against our policy to assert on everything.
(see vca_tcp_opt_set() for example).

So, for the time being, I think simple serialization using a pthread mutex
is the safest and simplest solution. A varnish lock would be overkill, as
this lock is really only relevant during VCA_Shutdown.

IMHO, the primary risk involved here is delayed VCA_Shutdown() due to
vca_acct() hanging in a syscall. Famous Last Words - sigh.

Fixes #2482