Investigate splitting generated bindings out of script

[jdm]

After #2101 lands, we have the notion of DOM method traits. We could conceivably move the bindings code out of script and thereby improve compilation time if the bindings code only used these traits. LTO might allow us to devirtualize the calls, thereby negating the performance hit of introducing virtual calls into one of the hottest paths. This doesn't solve the problem of passing trait objects to and from C code via SpiderMonkey, of course; it suggests a lot of unsafe casting :/