Performance impact of --allow-redefinition-new

[ilevkivskyi]

I just noticed that on self-check --allow-redefinition-new makes it ~10% slower. Probably because we infer more union types. I think we should fix (or at least investigate) this before renaming --allow-redefinition-new to --allow-redefinition.

cc @JukkaL