perf: Optimize NULL handling in array_slice

[neilconway]

Which issue does this PR close?

• Closes Optimize NULL handling in array_slice #21481.

Rationale for this change

array_slice does a per-row NULL checking (in four different arrays!). It would be faster to take the union of the four input NULL buffers via NullBuffer::union.

Benchmarks (Arm64):

  - List(Int64), array args:                  60.98ms -> 58.35ms  (-4.3%)
  - List(Int64), array args, no stride:       28.19ms -> 25.92ms  (-8.0%)
  - List(Int64), scalar args, no stride:      57.07ms -> 55.56ms  (-2.6%)
  - List(Int64), scalar args, stride=-2:      99.44ms -> 96.05ms  (-3.4%)
  - List(Int64), scalar args, stride=-1:     155.23ms -> 155.30ms (+0.0%)
  - List(Int64), scalar args, stride=1:       58.50ms -> 55.91ms  (-4.4%)
  - List(Int64), scalar args, stride=2:      151.45ms -> 146.83ms (-3.1%)
  - ListView(Int64), array args:              56.19ms -> 52.86ms  (-5.9%)
  - ListView(Int64), array args, no stride:   28.53ms -> 24.35ms  (-14.7%)
  - ListView(Int64), scalar args, no stride:  58.65ms -> 58.34ms  (-0.5%)
  - ListView(Int64), scalar args, stride=-2:  93.85ms -> 91.59ms  (-2.4%)
  - ListView(Int64), scalar args, stride=-1: 149.68ms -> 149.06ms (-0.4%)
  - ListView(Int64), scalar args, stride=1:   59.53ms -> 58.90ms  (-1.1%)
  - ListView(Int64), scalar args, stride=2:  143.07ms -> 139.55ms (-2.5%)

What changes are included in this PR?

Are these changes tested?

Yes.

Are there any user-facing changes?

No.