As a consequence, using dynamically a version of GMP which
does not match the compile time version is no longer an error
(this patch makes this a warming).
The code is shorter, but we break the MPZ API boundary; this also
allows to put allocating pointer_to_array out of the loop, and get
slightly better performances.

Yu Yichao indicated the thread-safety issues and the skeleton of
a solution, and explained the need to keep references to BigInt
before ccall, and suggested to add a field at the end of MPZ.
He also pointed at the GC un-safety of the previous solution
due to the fact that finalizers can call arbitrary code. Scott
Jones suggested to disable the GC to address this specific problem.