Optimize KeyFormatter on cache hits

[moberegger]

Builds on top of #597. This PR optimizes the KeyFormatter so that it doesn't lock the mutex on a cache hit by implementing the check-lock-check pattern, which gives us a fast non-locking path for the most likely case. This was inspired by how Concurrent::Map is implemented.

The mutex will only be locked when there is a cache miss, and while locked it will first attempt another read to account for the situation where another thread has already computed the value.

This small change results in a significant improvement on cache hits when running Jbuilder::KeyFormatter#format.

ruby 3.4.7 (2025-10-08 revision 7a5688e2a2) +YJIT +PRISM [arm64-darwin24]
Warming up --------------------------------------
              before   852.760k i/100ms
               after     1.566M i/100ms
Calculating -------------------------------------
              before     10.275M (± 4.7%) i/s   (97.32 ns/i) -     52.018M in   5.077786s
               after     23.927M (± 1.7%) i/s   (41.79 ns/i) -    120.613M in   5.042395s

Comparison:
              before: 10274977.5 i/s
               after: 23926727.6 i/s - 2.33x  faster

When using key formatting, this is one of the hotter code paths in jbuilder. The following benchmark

json = Jbuilder.new(key_formatter: Jbuilder::KeyFormatter.new(:downcase))

Benchmark.ips do |x|
  x.report('before') do |n|
    n.times { json.set! :Foo, :bar }
  end
  x.report('after') do |n|
    n.times { json.set! :Foo, :bar }
  end

  x.hold! 'temp_ips'
  x.compare!(order: :baseline)
end

Shows a 1.29x improvement

ruby 3.4.7 (2025-10-08 revision 7a5688e2a2) +YJIT +PRISM [arm64-darwin24]
Warming up --------------------------------------
              before   416.819k i/100ms
               after   521.578k i/100ms
Calculating -------------------------------------
              before      4.641M (± 2.9%) i/s  (215.45 ns/i) -     23.342M in   5.034002s
               after      5.993M (± 3.3%) i/s  (166.86 ns/i) -     30.252M in   5.054337s

Comparison:
              before:  4641344.3 i/s
               after:  5993038.9 i/s - 1.29x  faster