These don't implement anything yet that the old syscalls didn't
support, but they give us a place to pass flags and make it easier to
share user code between xv6 and Linux.

This fixes the syscall path so it no longer clobbers %r9, which means
we can use it as the 6th function argument, following the AMD64 ABI.
We still need a scratch register, but we now get that by stashing away
%rax (the syscall number) in a per-CPU variable, making it available
for use.

Bet you saw that coming.

It makes no sense, since it doesn't declare any avars, but is
obviously capable of having physical sharing.  This would only make
sense if we didn't *also* declare ascopes in individual system calls.

Renaming this "pte" more closely represents what's being shared.
Also, previously we were inconsistent about including the vmap or not;
now we use the same format everywhere.

E.g. make RUN="avar on; ls; avar off; halt"

QEMU copies the command line into the bss, which makes saving it a
slightly annoying process.

The only locking is "point-to-point" between condvar sleep in one proc
and condvar wakeup in another.

half-broken implementation of syscall/sysret; exec seems to work
fine with syscall/sysret now.

A change in any kernel source may change the syscall definitions, so
those need to be regenerated, but previously this would cause
everything else to be likewise regenerated.  Avoid this by only
updating the generated file if its contents actually change.

This also makes these rules better behaved when there are errors,
since previously they would leave partially-generated files that would
confuse the next make run.