Features/reduce gc pressure, massive speed improvements in benchmark (if Bullet.enable)

[riffraff]

This branch makes a few changes to reduce object allocation, uses a Set in Registry::Base so that we can avoid uniq calls all the time, moves flatten in a different position so it's not processing all the registry all the time and uses a direct hash access instead of iterating on a hash. Plus a couple spurious fixes.

I also merged and updated performance/benchmark.rb so it enables bullet and uses profiletools.rb.
From my investigation it resulted that bullet code was not being invoked otherwise.
All tests still passing on 1.8 and 1.9, (osx, linux).

Result for 1.8.7, standard code benchmark (profiling enabled)

 Start benchmarking...
                                                                             user     system      total        real
 Querying & Iterating 1000 Posts with 10000 Comments and 100 Users     405.980000   0.440000 406.420000 (406.522312)

1.8.7, my branch

 Start benchmarking...
                                                                             user     system      total        real
 Querying & Iterating 1000 Posts with 10000 Comments and 100 Users       3.290000   0.030000   3.320000 (  3.369753)

On ruby 1.9.3, no perftools (no 1.9.2 at hand sorry), standard code

Start benchmarking...
                                                                        user     system      total        real
Querying & Iterating 1000 Posts with 10000 Comments and 100 Users     ^[[B305.490000   1.630000 307.120000 (306.878098)

my branch

Start benchmarking...
                                                                            user     system      total        real
Querying & Iterating 1000 Posts with 10000 Comments and 100 Users       3.250000   0.050000   3.300000 (  3.300147)

I don't understand why some things were done in such a way, so this may break something else.

[flyerhzm]

Thanks man, I will try it this weekend.