zipper_algebra: pass a more reasonable child mask

[marcin-rzeznicki]

This leads to a few percent speed-up in the benchmarks. For instance, for symmetric difference simulation, before this we had:

Timer precision: 33 ns
ops               fastest       │ slowest       │ median        │ mean          │ samples │ iters
╰─ xor                          │               │               │               │         │
├─ base_2b     71.11 ms      │ 72.15 ms      │ 71.48 ms      │ 71.55 ms      │ 5       │ 50
├─ base_4b     838.6 ms      │ 858.7 ms      │ 840.8 ms      │ 844.1 ms      │ 5       │ 50
├─ zipper2_2b  308.1 ms      │ 312.8 ms      │ 310.3 ms      │ 310.2 ms      │ 5       │ 50
├─ zipper2_4b  2.625 s       │ 2.65 s        │ 2.635 s       │ 2.636 s       │ 5       │ 50
├─ zipper4_2b  158.1 ms      │ 159.3 ms      │ 158.7 ms      │ 158.7 ms      │ 5       │ 50
╰─ zipper4_4b  1.603 s       │ 1.617 s       │ 1.615 s       │ 1.612 s       │ 5       │ 50

and now we have

Timer precision: 31 ns
ops               fastest       │ slowest       │ median        │ mean          │ samples │ iters
╰─ xor                          │               │               │               │         │
├─ base_2b     75.49 ms      │ 75.92 ms      │ 75.67 ms      │ 75.67 ms      │ 5       │ 50
├─ base_4b     850.8 ms      │ 868.2 ms      │ 857.7 ms      │ 858.6 ms      │ 5       │ 50
├─ zipper2_2b  315.6 ms      │ 320.1 ms      │ 317.7 ms      │ 317.7 ms      │ 5       │ 50
├─ zipper2_4b  2.147 s       │ 2.194 s       │ 2.15 s        │ 2.167 s       │ 5       │ 50
├─ zipper4_2b  155.8 ms      │ 158.4 ms      │ 155.9 ms      │ 156.5 ms      │ 5       │ 50
╰─ zipper4_4b  1.285 s       │ 1.312 s       │ 1.286 s       │ 1.296 s       │ 5       │ 50

so there is another quite considerable speed-up (especially for longer paths)

PS. The actual changes are small and contained in d2d3abc