Adding "-mfloat-abi=hard" to LDFLAGS fixes the issue for me, and also allows $(BUILDPYTHON) to be built correctly in addition to $(PGEN).
So I guess the resolution to this issue is "works for me" (with setting CC or LDFLAGS properly for cross-compilation being the resolution/workaround).
Does it make sense to create a new bug "Do not build pgen when it's not going to be used" as follow-up to this discussion (with a patch similar to the one in http://bugs.python.org/msg270304)? If so, I'll create one. Or maybe there should be a generic configure flag "do not run any generators" (like https://bugs.python.org/issue26662#msg270162, but including not running pgen), with the flag only issuing warnings instead of failing.