Explore differences in LFRic KGOs

[sergisiso]

Without exact match is difficult to guarantee that we haven't introduced small errors

Some ideas why they could be differences:

• It is compiler-dependent and the baseline is gcc
• It is number-of-ranks dependent but we just have one for all tests
• We may need the config REPRODUCIBLE_REDUCTIONS=true
• We may need the 'math_uniform' in gpu run