Fix in-register representation of fdiv.s/fsqrt.s result

We were zero-extending it, which is a double-precision zero in the recoded
format.  So, when spilled and reloaded with fsd/fld, the original value
was destroyed.  Instead, set the MSBs so that it represents sNaN.  When
spilled, the single-precision number will be preserved as the NaN payload.