Fold integer endpoint upweighting into weights=

[tybug]

This makes st.integers map to a single draw_integer call in our IR, improving efficiency for shrinking, and possibly generation due to e.g. example copying (generate_mutations_from). Previously discussed at #4101 (comment) and other intermittent places.

The new interface is:

    @abc.abstractmethod
    def draw_integer(
        self,
        min_value: Optional[int] = None,
        max_value: Optional[int] = None,
        *,
        # previously:
        # weights: Optional[Sequence[float]] = None,
        weights: Optional[dict[int, float]] = None,
        shrink_towards: int = 0,
        forced: Optional[int] = None,
    ) -> int:

I've made several practical compromises here to ease implementation:

• sum(p) must be strictly less than 1. In other words, weights cannot be a total interval. This allows the forcing implementation to force into the remaining probability mass, which is not possible if it is 0.
• The remaining probability mass - 1 - sum(p) - gets distributed over all values, not just the unmapped ones. I think this is fine for current usages, but something to improve in the future.

This approach is actually not my preferred api interface. I would like to have weights: dict[IntervalSet, float] which maps intervals to p. This would make it possible to express things like "downweight everything below zero". Unfortunately, I had difficulty getting this approach to work with shrinking. My WIP branch is at https://github.com/HypothesisWorks/hypothesis/compare/master...tybug:hypothesis:integer-weights?expand=1, but I don't expect to come back to it in the near future.

---

@pschanely I think you ignore weights in hypothesis-crosshair, so I don't expect this to change anything for you, but I thought you'd appreciate a heads up for interface changes regardless 🙂

[tybug]

I got tired of seeing this flake (which, as I discovered, happens because the number of bits of the two integers were not equal, failing a check in redistribute_block_pairs), so I rewrote it into redistribute_integer_pairs.

[Zac-HD]

Looks great - thanks again @tybug!

[Zac-HD]

This approach is actually not my preferred api interface. I would like to have weights: dict[IntervalSet, float] which maps intervals to p. This would make it possible to express things like "downweight everything below zero". Unfortunately, I had difficulty getting this approach to work with shrinking.

Oooh, that would be very nice. I suspect this will be trivial after we drop the bytestring representation and pretty annoying until then, so probably we just have to work through all the other cleanups first 😿