Changing the `Binary` instance for `Double` and `Float`

I ended up creating primitive-serial for this exact issue. All it does is pull representations out of memory using peek, then swap over bytes as needed depending on CPU endianness.