x86/x86_64: `lock xor` instead of `cmpxchg` loop ({8,16,32}-bit atomics on x86, but additionally 64-bit atomics on x86_64)
- MSP430: `xor` instead of disabling interrupts ({8,16}-bit atomics)

Note: On x86/x86_64, the use of either function should not usually
affect the generated code, because LLVM can properly optimize the case
where the result is unused.

# Examples

```
use portable_atomic::{AtomicI16, Ordering};

let foo = AtomicI16::new(0b101101);
foo.xor(0b110011, Ordering::SeqCst);
assert_eq!(foo.load(Ordering::SeqCst), 0b011110);
```