Merging $ref support had a 10x performance hit

[Julian]

Merging #37 / #41 put performance on the perftest down the tubes.

⊙  git checkout -q a8656b8d75944a0029376a0ed78f91c0cb7c886f && ./perftest
Validated 1000 times. Finished in 11.2677850723 seconds.

⊙  git checkout -q 0e87ffafc8c2bd && ./perftest
Validated 1000 times. Finished in 1.6240568161 seconds.

I have a slight feeling that this may be due to a poorly written performance test (or at least one that doesn't properly reflect real world use) but this definitely needs to be investigated.

[Julian]

Never mind. This seems to just be because of all the "ref" : "#"s in the meta schema, which is now doing a whole lot more validation now that ref actually works.

Not that that means that we shouldn't have a less shitty perftest, but this is correct.