Better internal implementation for 'filtered choice'

[Zac-HD]

As discussed in the shrinking guide, we have an internal trick that can make sampling from a filtered list much more efficient. It would be great to translate sampled_from(...).filter(...) into this form internally!

---

On the implementation side, the most elegant way to do this is probably to extend FilteredStrategy.do_draw with special cases for strategies that are choosing from a fixed number of elements, i.e.:

1. Choose a random element, and if it satisfies the filter return it. Try this three times, because rejection sampling has low overhead.
2. [new bit] If the wrapped strategy is a sampled_from of reasonable size (e.g. <1k). This could extend to booleans() and just() in future
   • We're going to calculate a list of allowed values, pick of of them, and write that element's index in the full list to the buffer (see shrinking guide for details)
   • Start by choosing an index k into the filtered list. We don't know how many there are, so just assume that it's unfiltered-1 (as we've failed a random draw).
   • Create the filtered list until we get to the kth element. This is up to twice as fast on average as creating the full list - less if very few elements are allowed, but probably more when we're shrinking.
   • If there are less than k allowed elements, pick a new k which is a valid index. If there are no valid elements, fall out of this condition - it'll be noted and marked invalid as if rejection sampling failed and we didn't try this special case.
   • Write the index of the chosen element in the unfiltered list to the buffer, and return the element.
3. Give up, because if there is a valid element we can't find it.

Where this gets a bit hairy is in the interaction with LaxyStrategy and delayed validation, so we'll probably need to fiddle with the delegation of the do_draw or filter methods a bit to have it work. That's why we have this write-up of the discussion from #1862.

[Zac-HD]

Per discussion on #1887, we should ensure that this is also implemented for unique collections via lists(..., unique=True).