Improve arrays() performance with unique-sampled-list tricks

[Zac-HD]

UniqueSampledListStrategy has significantly better performance than the approach currently taken by arrays() when choosing the indices at which to insert elements. We can and should leverage that, and use similar tricks for selecting unique elements for eight-bit dtypes where high collisions are possible. This also implies that we should usually choose an element and then the index to avoid wasting entropy if the element is rejected.

Note that the same considerations apply to #3065 as well as hypothesis.extra.numpy; the relevant code is identical.