Make use of the full range of AUX matrices

The gist of it is that we have this signature `mr2d_malloc(Int n)`,
bounding the internal `malloc` to `Int`, which is Scalapack's indexing type.
This is not how `malloc` is defined and means that maximum AUX matrix
size is artifically limited on 64-bit systems to 2GB.

This PR operates this transformation: `mr2d_malloc(Int n) -> mr2d_malloc(size_t n)`,
yet ensures `Int -> size_t` conversion does not involve Int-negative values
and ensures that no 64-bit values are passed to `malloc` on 32-bit systems.

The main advantage of this on 64-bit system allows the use of
the **full** _signed_ 32-bit indexing range instead of `range / element size`.
E.g. the max AUX matrix size is now 16GB instead of 2GB previously.

With this, "standard" 32-bit Scalapack and Blas/Lapack can still be used
in programs like GPAW.

Details (and full commit history) are in
Reference-ScaLAPACK/scalapack#85
which does not seem to receive much attention.

This patch is a way to provide that feature to conda users in the meantime.