libio mmap changes

[Roland McGrath]

> Well, this is the cleanest solution but...  Function tables in DSOs are 
> not cheap.  I cannot say in the moment whether there is a better method 
> but I want to mention this here for the benefit of all.  If you have to 
> design some code in future avoiding function tables is advised.

Yes, this had occurred to me.  At least all the symbols are local/hidden so
it should only be RELATIVE relocs, no?  And since you added one for mmap in
the first place...  If it is kosher to add a flag bit somewhere, then the
other obvious approach is to have the old single file jump table and have
each operation test a flag bit--another test and branch on every callback
vs the startup relocation of a couple dozen words (and ones that can never
have conflicts if prelinked).