Update cfloat.h to fix subnormal casting and rounding

- Rewrites the subnormal and rounding code to fix a bug which resulted in
very small inputs being transformed into much larger destination values.
- Adds an explicit flush-to-zero control for the cast operator.

Signed-off-by: Jerry Ge <jerry.ge@arm.com>
Change-Id: Ia165d026f7764189c4f1704b7407493f55118597