Build using --with-gmp and shared libraries

[Tobias.Schlueter@Physik.Uni-Muenchen.DE]

Quoting Paolo Bonzini <paolo.bonzini@lu.unisi.ch>:
> As a side note, making an in-tree mpfr possible is something easy to do,
> that I'd endorse, and that I could help with (unlike moving libgcc to
> the toplevel ;-).

It's something that we want, together with an in-tree gmp (mpfr depends on gmp
after all).  There were some past conversations about this, and I remember
having been told that at least one gcc heavyweight had failed doing this due to
some weirdness either in gmp or mpfr.  I'm cc'ing Paul Brook as I think that I
got that information from him.

- Tobi