regression introduced with !195

!195 (merged) does not really work as it sets a variable in a macro that might (and is) redefined when find_package is invoked.

Example: from lb-dev LHCb we first look for LHCb without the constrained version, then LHCb looks for Gaudi constraining the version, but that leaks the value of _require_exact_version to the outer scope (it's not really outer as they are macros) and we get an error.