Improve SIMD warnings - instructions sets are not incremental
Nandan has reported crashes he could not understand where the output merely contained a minor recommendation:
Compiled SIMD: AVX_128_FMA, but for this host/run AVX2_128 might be better (see log).
This specific crash happens because AVX2 is not a superset of AVX_128_FMA (which happens to contain AMD-specific instructions.
However, apart from a specific error message, in particular our MPI code makes assumptions that SIMD instruction sets are incremental and just tries to identify the "lowest" and "highest" level supported, which in general is not true. I believe we need to change this to always provide a short warning when SIMD instruction sets are mismatching (in particular when nodes are different in an MPI job), and only make explicit exceptions to be quiet in cases where we know we should NOT warn.