rtld build changes

rtld build changes

[Roland McGrath]

I have committed an updated version of the changes I posted previously to
reorganize the way ld.so is built.  Along with this are several cleanups
and optimizations that this enables, and we may want to do more now that we
have the facility.  These changes are all working fine for me, but I have
only tried builds on x86, ppc, and x86-64, and not done much testing beyond
make check.  So be on the lookout for sudden wierdness and let me know.

So far two things of interest were made possible by building ld.so the new way:

ld.so always uses its own private errno variable (to which all access is
already serialized), and all accesses to it use GOTOFF relocs.  Between
this and some other optimization/cleanups, ld.so has gone from 16 relocs to 9
and that's pretty much as low as it can go.

When libc is using __thread, ld.so still does not.  You can now have one
ld.so installed capable of loading a libc binary built to use __thread or
one built without it.  This also eliminates some startup overhead
(including a system call) from ld.so.