Fix test for kernel-generated signals when re-raising.

Testing si_pid to determine if a signal was sent by another process with a
signal like kill() isn't always right.  The si_pid field is valid only when the
signal was actually sent by a process (either with a syscall like kill() or, for
SIGCHLD, by exiting).  In particular, on Linux it overlaps with the si_addr
field, so an actual crash will often have nonzero si_pid.

I think what we actually want to do is test si_code; the Linux kernel reserves
positive values for kernel-generated signals (and this is also specified by XSI,
more or less — see the application usage section of
http://pubs.opengroup.org/onlinepubs/9699919799/basedefs/signal.h.html ).

I'm not sure if my patch does the right thing for the SysRq SIGABRT case; if
someone has a pointer to the kernel code that generates it, that would help.

[Ted Mielczarek]

Lei, can you take a look at this?