Produce better code from maskW

```
complement (m-1) `xor` m == -m `xor` m
```

Latter produces slightly better assembly.

```asm
	movq	%rdi, %rax
	negq	%rax
	xorq	%rdi, %rax
```

Indeed if we give

```cpp
uint64_t f(uint64_t m_aQU2) {
  return ((m_aQU2 - 1) ^ 18446744073709551615UL) ^ m_aQU2;
}
```

to clang and compile with optimisation, that's the ASM it spits out.
Translating it back to Haskell gives the solution from this branch.
Without same optimisation, even in Haskell it saves an instruction.

```hs
maskW :: Word -> Word
maskW m = complement (m-1) `xor` m

{-
	movq %rax,%rbx
	xorq $-1,%rbx
	decq %rax
	xorq %rbx,%rax
-}
```

```hs
maskMine :: Word -> Word
maskMine m = (-m) `xor` m

{-
	movq %rax,%rbx
	negq %rbx
	xorq %rax,%rbx
-}
```

Author: Fuuzetsu

Data/IntMap/Internal.hs

@@ -1117,7 +1117,7 @@ withoutKeys t1@(Bin p1 m1 _ _) (IntSet.Tip p2 bm2) =
     let minbit = bitmapOf p1
         lt_minbit = minbit - 1
         maxbit = bitmapOf (p1 .|. (m1 .|. (m1 - 1)))
-        gt_maxbit = maxbit `xor` complement (maxbit - 1)
+        gt_maxbit = (-maxbit) `xor` maxbit
     -- TODO(wrengr): should we manually inline/unroll 'updatePrefix'
     -- and 'withoutBM' here, in order to avoid redundant case analyses?
     in updatePrefix p2 t1 $ withoutBM (bm2 .|. lt_minbit .|. gt_maxbit)
@@ -3336,7 +3336,7 @@ mask i m
 -- bit @m@.
 maskW :: Nat -> Nat -> Prefix
 maskW i m
-  = intFromNat (i .&. (complement (m-1) `xor` m))
+  = intFromNat (i .&. ((-m) `xor` m))
 {-# INLINE maskW #-}
 
 -- | Does the left switching bit specify a shorter prefix?