Use set data structure to speed up circular reference checks on large/deeply nested objects

[alindeman]

This is a refactor, so no updates to tests.

This also could be cherry picked cleanly into 3-1-stable.

[josevalim]

Do you have some simple benchmark that shows that using set is really faster? Thanks!

[alindeman]

Given (I've seen objects even "bigger" than this need to be encoded for certain special cases):

big_object = 500.times.map { |i| Hash[('a'..'z').map { |j| [j, j] }] }

ruby-1.9.2-p180 :016 > Benchmark.bm do |x|
ruby-1.9.2-p180 :017 >     x.report('array') { 100.times { ActiveSupport::JSON.encode(big_object) } }
ruby-1.9.2-p180 :018?>   end
      user     system      total        real
array 28.600000   0.100000  28.700000 ( 28.703123)

ruby-1.9.2-p180 :011 > Benchmark.bm do |x|
ruby-1.9.2-p180 :012 >     x.report('set') { 100.times { ActiveSupport::JSON.encode(big_object) } }
ruby-1.9.2-p180 :013?>   end
      user     system      total        real
set 27.440000   0.110000  27.550000 ( 27.546195)

While the numbers varied slightly over multiple runs, set always beat array by a significant enough margin for me to think this small patch is worth applying.