added g-i-omkl-2020a toolchains

[mboisson]

This is to split the cuda part from the PR #10954

@Micket

[mboisson]

This PR puts MKL at the SYSTEM level. This means that toolchains such as gmkl and iimkl, which include MKL but does not include MPI, are possible.

This allows - for example - building R with MKL, without having to place it at the MPI level. In fact, a gcccoremkl is even possible (although not include in this PR), allowing to install software which depend on MKL, but not on vectorization at the core level. This avoids having to recompile those pieces of software multiple times.

[boegel]

@mboisson This needs a sync with develop now #11036 is merged...

[mboisson]

@boegel done

[mboisson]

This is probably no longer relevant