For instance, try entering just a long word as the regex, e.g. "helloworld":
%state0.goto.1.0.cmp_mask = icmp eq <10 x i8> %state0.rhs, < i8 104, i8 101, i8 108, i8 108, i8 111, i8 119, i8 111, i8 114, i8 108, i8 100 > ; helloworldHigh performance regex engines do use SIMD in various places. Others have mentioned Hyperscan, which is undoubtedly a showcase of the most sophisticated SIMD techniques when it comes to regex and substring searching. But even something like RE2 uses vectorized code in places, although indirectly.
But, in all instances I can think of, using SIMD is a matter of identifying some optimization opportunity in regex matching that doesn't generalize to handling all cases. At some point, to handle some cases, you'll need that "one character at a time" loop. (Whether it's backtracking or FSM based.) So visualizations like these are still rather helpful, and at the very least, give a basic conceptual understanding of what the most general kind of FSM-based regex might look like internally. (Although, most general purpose regex engines don't actually utilize all of the transformations presented in the OP.)
Regex is used all over the place, so it's certainly worthwhile to speed things up. And huge FST's are also used in language tech, e.g. to analyze all frillions of the forms of Finnish verbs into all their ambiguous readings with tags and dictionary forms: https://beta.apertium.org/index.eng.html?choice=fin&qA=Taite... The language supported by that FST is basically infinite due to derivations and compounding, but the zipped binary fits on a handful of floppies.