libc requires this to use `long double` which is sometimes the same as
f128, sometimes not.

Also for an unknown reason, aarch64 is getting an invalid result for the
`@mulAdd` behavior test for f128. See #9900.