perf: skip redundant downcase in Currency.find! on cache hits

[cribbles]

Summary

Optimize Currency.find! (aliased from Currency.new) to skip to_s.downcase on cache hits. On the first resolution of a currency string (e.g. "USD"), the result is stored under both the original-case and downcased keys in @@loaded_currencies. Subsequent lookups hit the fast path immediately without allocating a downcased string or checking the mutex.

Stacks on #520.

Benchmark (2M iterations of Money.new)

require "money"
require "benchmark"

n = 2_000_000
c = Money::Currency.find!("USD")

Benchmark.bm do |x|
  x.report("string") { n.times { Money.new(100, "USD") } }
  x.report("currency") { n.times { Money.new(100, c) } }
end

main

              user     system      total        real
string    1.801176   0.008822   1.809998 (  1.942647)
currency  1.305507   0.004958   1.310465 (  1.423729)

#520 (base branch)

              user     system      total        real
string    1.735361   0.005490   1.740851 (  1.860185)
currency  1.237422   0.003729   1.241151 (  1.351668)

This branch

              user     system      total        real
string    1.618875   0.007278   1.626153 (  1.797001)
currency  1.200702   0.005228   1.205930 (  1.378954)

Summary

                    main → #520      #520 → this     main → this
string              ~4%              ~7%              ~10%
currency            ~5%              ~3%              ~8%

[elfassy]

I think we're better off improving the find! method (already cached) than adding another layer of caching. Ex #519 (comment). The only difference is that this PR is not downcasing the currency string. This means the cached hash could be double the size if people use both lower and upper case, making each search slower and use more memory. You can account for the larger hash by looping through every currency in both upcase and lowercase a few times in your benchmark