r19412: some rather strange looking changes to talloc that gain us about 50%

r19412: some rather strange looking changes to talloc that gain us about 50%
in the talloc benchmark. These changes were driven by some cachegrind
profiles, with the biggest improvements coming from inlining some
functions.

I don't think it would be a good idea to start spreading inline and
likely()/unlikely() in other parts of Samba, as the benefit in most
code will be very small, but talloc() is such a speed critical part of
Samba that I think these changes are worthwhile