Use ConjectureData.for_choices

[tybug]

Depends on #4217 (so test_draw_to_overrun finishes in a sane amount of time).

Here are three proposals for a replacement for ConjectureData.for_buffer on the typed choice sequence:

(1) pass the value of each choice:

    @classmethod
    def for_choices(
        cls,
        choices: Sequence[NodeTemplate | IRType],
        ...
    ):
        ...

(2) pass the value of each choice and its complexity index:

    @classmethod
    def for_choices(
        cls,
        choices: Sequence[NodeTemplate | tuple[IRType, int]],
        ...
    ):
        ...

(3) pass the value of each choice and its kwargs, noting that the complexity of a choice can be derived from its value + kwargs:

    @classmethod
    def for_nodes(
        cls,
        nodes: Sequence[NodeTemplate | IRNode],
        ...
    ):
        ...

I am proposing we use (1) in this pull. The benefits of this is mainly a simplified and unified api. It also makes constructing a ConjectureData from a database entry possible, where we do not have choice complexity available. However, I am leaning towards encoding complexity alongside db values anyway, for more precise sorting, so this may be a nonissue.

The downside of (1) is that we cannot use choice complexity during drawing, in particular misalignment. This leads to 129da4b, which fills misalignments with the zero-index choice instead of using complexity as an intermediary. I think using NodeTemplate to opt-in to controlling this misalignment behavior in the shrinker could be a more clean path forward. In the mean time, the default and only behavior for correcting misalignments would be filling with the zero-index choice.

The shrinking benchmark is neutral:

@DRMacIver if you have any thoughts on this from a shrinking perspective I'd definitely be curious to hear them.

[Zac-HD]

LGTM overall, feel free to merge once the comments below are handled to your satisfaction.

[Zac-HD]

I'd be inclined to assert isinstance(...) rather than ignoring, since this is unlikely to be a noticable perf hit.

[tybug]

assert isinstance(ChoiceT) is impossible and assert not isinstance(NodeTemplate) feels bad - I went for a slightly more roundabout type checking solution.