getAddrInfo and getNameInfo lose their errors on Windows #636

On Windows, every failure of getAddrInfo and getNameInfo is raised as the same userError, and what failed and why is lost.

Reproduce

import Network.Socket
import qualified Control.Exception as E
import System.IO.Error (ioeGetErrorType)

main = do
    r <- E.try (getAddrInfo Nothing (Just "no-such-host.invalid") (Just "80"))
    case r of
        Left e   -> print (e :: E.IOException) >> print (ioeGetErrorType e)
        Right as -> print (map addrAddress (as :: [AddrInfo]))

On Windows 11 with GHC 9.12.4 and master:

user error (Network.Socket.gai_strerror not supported: 11001)
user error

On macOS:

Network.Socket.getAddrInfo (called with preferred socket type/protocol: Nothing, host name: "no-such-host.invalid", service name: "80"): does not exist (nodename nor servname provided, or not known)
does not exist

Cause

configure.ac checks with AC_CHECK_FUNCS([gai_strerror]), which links against a function of that name. In mingw-w64, gai_strerror is a macro for gai_strerrorA, so the check fails (config.log: undefined symbol: gai_strerror) and HAVE_GAI_STRERROR is not defined. gai_strerror in Network/Socket/Info.hsc then throws userError "... not supported" in place of the error it was asked to describe.

Expected

The same as on POSIX: NoSuchThing with the location, the arguments and a description of the code (11001 is WSAHOST_NOT_FOUND).

The codes there are Winsock codes, so getWSErrorDescr in cbits/winSockErr.c can describe them; it already has WSAHOST_NOT_FOUND, WSATRY_AGAIN, WSANO_RECOVERY and WSANO_DATA. That also avoids gai_strerror on Windows, which writes to a static buffer and is not thread-safe. The code could also be kept in ioe_errno, as it is for other Winsock errors (#631).

Related: #635.