Issue Description:

Consider a parallel `forEach`. We know the closure doesn't escape the call, but it needs to be copied to all threads. Even if we `unsafeBitCast` the type of the closure away, there's no way to call it without inducing a retain/release pair, as you can see here: https://swift.godbolt.org/z/igFqV9

Of course I realize the type aliasing is probably illegal, but it serves to illustrate the point. The function pointer in the closure uses a calling convention that matches neither @convention(c) nor @convention(thin), so there's no way to invoke it and manually pass the context pointer without generating ARC traffic. We really need some kind of callable UnmanagedFunction<F> type or @convention(rawSwift) to represent the raw function pointer in the closure.