Check subject UTF8 validity just once for String#gsub, #scan, #split

[HertzDevil]

As of #13313, regex matching uses neither NO_UTF_CHECK nor MATCH_INVALID_UTF, which means the subject string is checked for UTF-8 validity once per match. String#gsub, #scan, and #split are special in that the regex doesn't change and is expected to be matched many times, so for these methods we should only need to do this check exactly once; if the first match does succeed, we know that the subject string is validly encoded and we can apply NO_UTF_CHECK in subsequent matches, regardless of whether other optimizations are in place.

Benchmarks:

require "benchmark"

class String
  def scan_new(pattern : Regex, &) : self
    # this PR's implementation
  end
end

data = "a" * 10000
count = 0

Benchmark.ips do |b|
  b.report("old") do
    count = 0
    data.scan(/a/) { count += 1 }
  end

  b.report("new") do
    count = 0
    data.scan_new(/a/) { count += 1 }
  end
end

old  72.43  ( 13.81ms) (± 0.66%)  313kB/op  38.50× slower
new   2.79k (358.60µs) (± 0.94%)  313kB/op        fastest

If #13353 is merged then options |= :no_utf_check would apply to the parameters of the respective methods themselves.

[HertzDevil]

Another candidate is StringScanner whose subject string is supplied in the constructor and also never changes afterwards.

[straight-shoota]

This is a good improvement, but maybe (hopefully?) only termporary. It will become obsolete if/when #12020 gets implemented.