[COFF] Share the tail in delayimport symbol thunks

E.g. for x86_64, previously each symbol's thunk was 87 bytes. Now
there's a 12 byte thunk per symbol, plus a shared 83 byte tail
function.

This is similar to what both MS link.exe and GNU tools do for
delay imports.

Differential Revision: https://reviews.llvm.org/D64288

llvm-svn: 365823