Make earthdistance use version-0 calling convention if not USE_FLOAT8_BYVAL,

and version-1 if USE_FLOAT8_BYVAL.  This might seem a bit pointless, but the
idea is to have at least one regression test that will fail if we ever
accidentally break version-0 functions that return float8.  However, they're
already broken, or at least hopelessly unportable, in the USE_FLOAT8_BYVAL
case.

Per a recent suggestion from Greg Stark.