Fixed

• unicode casefolding (#58), due to divergence in casefolding between aho-corasick (ascii only) and regex (full unicode) regexes containing uppercase unicode literals would not be properly selected by the prefilter

Optimized

• replaced IndexSet<NoHash> by a bespoke IntSet, yielding ~10% throughput improvement and removing the dependencies on indexmap and nohash
• repetitions expansions into atoms (#53), repetitions would not be expanded leading to atoms of length 1, and unnecessary acceptance of regexes
• quality of atoms generated by concatenation (#53), allowing subsequent patterns to properly concatenate into atoms generated by concatenation

Other

• updated to edition 2024

• fix regex rewriting which didn't trigger properly in a significant number of cases: any repetition following an explicit character class would be left as-is, leading to memory use being cut down by ~60% (and #2 should have cut down memory use by 75% rather than 15-ish...) #29
• use an atom length of 2 for device parsers, improves performance by 40% at essentially no memory cost, don't touch ua and os as it reduces their performances by 510% instead #31 / #32

The performance fixes to regex-filtered definitely seem significant
enough to cut a release already, especially as they do impact
ua-parser quite dramatically (~30% performance improvement on the
bench script using DailyMotion's sample and the master regexes from
uap-core).