Finding the "lower case of a character" is immensely harder in face of unicode, because the table is way larger and there's no nice speed hacks by manipulating an index into the ASCII table. JFGI: "tolower performance unicode".
If you notice setting LC_ALL makes a difference in performance for you, you ought to be aware that now you're no longer comparing same capabilities.
I'm not sure how to constructively state that except for: One shouldn't compare ASCII and unicode "tolower" in performance comparisons, as you end up comparing apples and oranges (at best. More like apples and snails) - which I did.
If the author uses code that uses "tolower" in job interviews, i.e., evaluates candidates based on their input WRT case normalization - and he even writes ("This is Unicode-aware ...") - he should know that Unicode awareness is not ubiquitous, and comes at quite a cost. Knowing about unicode awareness, one would assume he'd be aware of not making a fair comparison.
The only one I see is the comparison between 'simple' and 'optimized'. But that's more about "what does a simple idiomatic solution look like" and what does an "optimized and possibly less simple" solution look like. That comparison isn't designed to be apples-to-apples in the way you're saying. The simple variant will use Unicode-aware casing in environments where that's the simple and natural thing to do.
IIRC, most of the 'optimized' programs are using ASCII casing. Python doesn't, but casing isn't even close to the bottleneck in that program.