Optimize ripper translator

[Earlopain]

Ref #3859 (comment). I was interested in how big the impact is and turns out it is a lot.

Creating state classes is pretty expensive. Since they are not modifiable, we can reuse them instead.

Benchmark script:

require "ripper"
require "prism"
require "benchmark/ips"

codes = Dir["**/*.rb"].map { File.read(it) }

Benchmark.ips do |x|
  x.config(time: 10)
  x.report("prism") { codes.each { Prism::Translation::Ripper.lex(it) } }
  x.report("ripper") { codes.each { Ripper.lex(it) } }

  x.compare!
end

Before:

ruby 4.0.0 (2025-12-25 revision 553f1675f3) +PRISM [x86_64-linux]
Warming up --------------------------------------
               prism     1.000 i/100ms
              ripper     1.000 i/100ms
Calculating -------------------------------------
               prism      0.293 (± 0.0%) i/s     (3.42 s/i) -      3.000 in  10.248348s
              ripper      0.633 (± 0.0%) i/s     (1.58 s/i) -      7.000 in  11.055687s

Comparison:
              ripper:        0.6 i/s
               prism:        0.3 i/s - 2.16x  slower

After

ruby 4.0.0 (2025-12-25 revision 553f1675f3) +PRISM [x86_64-linux]
Warming up --------------------------------------
               prism     1.000 i/100ms
              ripper     1.000 i/100ms
Calculating -------------------------------------
               prism      0.486 (± 0.0%) i/s     (2.06 s/i) -      5.000 in  10.280413s
              ripper      0.635 (± 0.0%) i/s     (1.58 s/i) -      7.000 in  11.027169s

Comparison:
              ripper:        0.6 i/s
               prism:        0.5 i/s - 1.31x  slower

[eregon]

Nice, thank you, it seems my performance guess was good then :)