MONGOID-4889 Optimize batch assignment of embedded documents (backport to 9.0)

[jamis]

(Backported to 9.0-stable)

When associating multiple embedded documents in a single assignment, the existing code was inefficiently looping over all records repeatedly, resulting in a significant performance impact. This PR attempts to minimize the number of loops needed to accomplish the same thing, while keeping the ordering of callbacks the same and otherwise attempting to not break backward compatibility.

To test the performance of this PR, the following benchmark (from MONGOID-4889) was used:

class Foo
  include Mongoid::Document
  embeds_many :bars
end

class Bar
  include Mongoid::Document
  embedded_in :foo
end

require 'benchmark'
array_1k = Array.new(1000) { Bar.new }
array_2k = Array.new(2000) { Bar.new }
array_3k = Array.new(3000) { Bar.new }
array_4k = Array.new(4000) { Bar.new }
array_5k = Array.new(5000) { Bar.new }

Benchmark.bm do |x|
  x.report('1k') { Foo.new.bars = array_1k }
  x.report('2k') { Foo.new.bars = array_2k }
  x.report('3k') { Foo.new.bars = array_3k }
  x.report('4k') { Foo.new.bars = array_4k }
  x.report('5k') { Foo.new.bars = array_5k }
end

With the previous implementation, the timings were abyssmal:

        user     system      total        real
1k  0.721604   0.002515   0.724119 (  0.724148)
2k  2.834939   0.008708   2.843647 (  2.843895)
3k  6.383648   0.017409   6.401057 (  6.401207)
4k 11.313247   0.035579  11.348826 ( 11.349941)
5k 17.895191   0.096859  17.992050 ( 18.048693)

Things look much better with the optimized implementation:

        user     system      total        real
1k  0.031262   0.000538   0.031800 (  0.031798)
2k  0.041094   0.000543   0.041637 (  0.041638)
3k  0.058909   0.000606   0.059515 (  0.059517)
4k  0.090872   0.000752   0.091624 (  0.091641)
5k  0.096289   0.001358   0.097647 (  0.097666)