Rewrite Windows `compat_fn` macro
This allows using most delay loaded functions before the init code initializes them. It also only preloads a select few functions, rather than all functions.
This is optimized for the common case where a function is used after already being loaded (or failed to load). The only change in codegen at the call site is to use an atomic load instead of a plain load, which should have negligible or no impact.
I've split the old `compat_fn` macro in two so as not to mix two different use cases. If/when Windows 7 support is dropped `compat_fn_optional` can be removed entirely.
r? rust-lang/libs
76822a2878