Soft float arithmetic should propagate NaN sign

[tgross35]

754 doesn't say anything about this but most hardware preserves NaN sign bit when one is propagated, so we should try to do the same.

[tgross35]

Relevant bit is here and similar

compiler-builtins/compiler-builtins/src/float/add.rs

Lines 35 to 41 in cb4a87f

	if a_abs > inf_rep {
	return F::from_bits(a_abs | quiet_bit);
	}
	// anything + NaN = qNaN
	if b_abs > inf_rep {
	return F::from_bits(b_abs | quiet_bit);
	}

, just need to or the sign bit back in.