Each core's gc thread free elements on the tofree list

Kernel is trickier because there are processes that don't call begin/end_epoch
Premature anyway, since it won't help until we have per-core process lists
But, we do garbage collect memory now

proc.c was using kfree, which places the two page stacks on the
single page free lists.

E.g.
  $ cat netif
  hw 52:54:00:12:34:56
  ip 10.0.2.15 nm 255.255.255.0 gw 10.0.2.2

Also disable annoying DHCP messages from kernel.

..I think we must use the SRAT to map physical memory to node IDs.

lwIP/net/e1000 seem to have a few nagging bugs, but httpd is able to
serve files.

lwIP crashes after a client connects and sh prints the prompt.

umemcpy is as safe as argptr.

scale-amd64 passes all "argint" syscall arguments via registers, so
it might be more convenient to pass them all to the sys_* functions
(instead of reading them via argint).