Make the use of "kernel32" more robust

[Delgan]

In response to a comment left on StackOverflow:

The project on GitHub should be updated to use kernel32 = ctypes.WinDLL('kernel32', use_last_error=True) and raise exceptions via raise ctypes.WinError(ctypes.get_last_error()). This (1) isolates it from other packages that use the global library loader ctypes.windll, which caches loaded libraries, which cache function pointers, and (2) reliably captures the thread's last error value in C immediately after the FFI call.