Replace i586 `ceil` and `floor` implementations with assembly to fix +/-0

[tgross35]

From rust-lang/libm#404 (comment):

i think for i586 it should just use frndint with setting the x87 CW (actually not slow) for floor/ceil. like: https://github.com/NetBSD/src/blob/20433927938987dd64c8f6aa46904b7aca3fa39e/lib/libm/arch/i387/s_ceil.S