Improve folds over bits for IntSet

[meooow25]

• Make foldr and foldl short-circuit instead of lazily accumulating thunks.
• Switch to a non-empty style to avoid unnecessary comparisons. This also helps GHC with arity-analysis (somehow), which greatly improves performance of CPS-style foldr and foldl.
• Change the bitwise operations used from bitmask = bm .&. -bm; bi = ctz bitmask; bm' = bm `xor` bitmask to bi = ctz bm; bm' = bm .&. (bm-1) which is slightly faster.

---

Benchmarks with GHC 9.10.1

Name                          Time - - - - - - - -    Allocated - - - - -
                                   A       B     %         A       B     %
dense.foldMap_elem            5.6 μs  5.3 μs   -4%      0 B     0 B
dense.foldMap_traverseSum     6.5 μs  4.5 μs  -31%    1.0 KB  1.0 KB   +0%
dense.foldl'_sum              5.6 μs  4.3 μs  -22%      0 B     0 B
dense.foldl_cpsOneShotSum      19 μs  5.7 μs  -70%    194 KB  2.5 KB  -98%
dense.foldl_cpsSum             20 μs  6.2 μs  -69%    162 KB  5.0 KB  -96%
dense.foldl_elem               12 μs  5.7 μs  -53%    162 KB  2.0 KB  -98%
dense.foldl_traverseSum        20 μs  5.9 μs  -70%    162 KB  5.0 KB  -96%
dense.foldr'_sum              5.8 μs  5.5 μs   -6%      0 B     0 B
dense.foldr_cpsOneShotSum      20 μs  5.6 μs  -72%    194 KB  2.5 KB  -98%
dense.foldr_cpsSum             21 μs  5.9 μs  -71%    162 KB  5.0 KB  -96%
dense.foldr_elem               12 μs  4.8 μs  -60%    162 KB  2.0 KB  -98%
dense.foldr_traverseSum        21 μs  4.9 μs  -76%    162 KB  5.0 KB  -96%
sparse.foldMap_elem            19 μs   18 μs   -5%      0 B     0 B
sparse.foldMap_traverseSum     21 μs   20 μs   -5%     64 KB   64 KB   +0%
sparse.foldl'_sum              16 μs   14 μs  -14%      0 B     0 B
sparse.foldl_cpsOneShotSum     58 μs   38 μs  -34%    319 KB  160 KB  -49%
sparse.foldl_cpsSum            58 μs   60 μs   +4%    287 KB  320 KB  +11%
sparse.foldl_elem              27 μs   29 μs   +6%    287 KB  128 KB  -55%
sparse.foldl_traverseSum       58 μs   55 μs   -5%    287 KB  320 KB  +11%
sparse.foldr'_sum              27 μs   24 μs  -10%      0 B     0 B
sparse.foldr_cpsOneShotSum     66 μs   29 μs  -55%    319 KB  160 KB  -49%
sparse.foldr_cpsSum            65 μs   50 μs  -22%    287 KB  320 KB  +11%
sparse.foldr_elem              37 μs   20 μs  -47%    287 KB  128 KB  -55%
sparse.foldr_traverseSum       65 μs   44 μs  -32%    287 KB  320 KB  +11%

---

Fixes #666

[meooow25]

Related GHC issue for a curious behavior I observed: GHC#25578 (also mentioned in the code).

[meooow25]

Okay, I'm pretty happy with this. Can't think of anything else that might be improved here.