Here's a patch, against the trunk, that simply amends load_float to ignore 
errno on underflow (while retaining it on overflow).  This fixes the problem on my machine.

The added tests may be a little severe.  There's no *good* reason why 
dumping and then loading a float on the same machine shouldn't produce 
exactly the same value, but there are likely many bad reasons.  I propose 
to checkin the change and watch the buildbots carefully;  if there are 
failures I'll weaken the tests to check for approximate equality instead 
of exact equality.