This patch adds linux native async mode support.

It is based on the same idea as Nick's patch.  The basic idea is that
listening for SIGCHLD signals, call waitpid in the handler until we
get all the pending waits out, and stuff the
results in a pipe.  This pipe is registered as the waitable file
in the event loop at the appropriate times.

I had to make a couple of target stack changes.  I had to replace
the  target_async_mask_value target variable by a proper target
method, which unfortunatelly makes me have to change remote.c
as well.

There are several target_can_async_p calls that are made *before*
the running target is pushed on the stack.  That problem wasn't
visible in "target async", because there you're testing against
a target that doesn't know how to run or attach anyway.
I fixed it by making the dummy target look for the default run
target like we do for the run and attach commands.
This is the part that I think I could split out of this patch,
but I though that showing it without showing this linux patch
wouldn't make much sense.

The patch still maintains support for non-async mode, which is
why we see caring to enable/disable in async handling in several
places.

With this patch on top of the pending posted async patches,
I only have regressions in define.exp.  Those are not caused
by this patch.  They are generic async problems (that have to
be fixed, but independently) that are visible in "target async" as well.

To enable/disable async mode, there's a new set/show command
"set linux-async 1" enables it.  I left it on in this patch, but
if/when this goes in, it should be turned off by default until
we fix at least the defines.exp problems.

To enable it from the command line, you can use
gdb -ex "set linux-async 1", of course.

(remote.c could take the same treatment and get rid
of "target async").

There are no regressions if async mode is disabled.

Comments ?  I'll be happy to clarify more.

-- 
Pedro Alves