heap_tuple_needs_freeze false positive

Hi,

I noticed that heap_tuple_needs_freeze might return true in cases where
the Xmax is leftover junk from somebody who set HEAP_XMAX_INVALID in the
far past without resetting the Xmax value itself to Invalid. I think
this is incorrect usage; the rule, I think, is that one shouldn't even
read Xmax at all unless HEAP_XMAX_INVALID is reset.

This might cause unnecessary acquisitions of the cleanup lock, if a
tuple is deemed freezable when in fact it isn't.

Suggested patch attached. I'd backpatch this as far as it applies
cleanly.

--
Álvaro Herrera <alvherre(at)alvh(dot)no-ip(dot)org>