Add initial coarse reduction pass for reducing alternatives

[DRMacIver]

This adds an initial phase to shrinking that is allowed to make changes that would be bad to make as part of the main shrink pass, with the main goal of producing better results for one_of.

This isn't as good a pass for this as I wanted to add. In particular it results in no new guarantees about shrinking, because we can't iterate this pass to a fixed point along with the others, but it should start the shrinking off from something that is much more likely to already be the minimal usable branch for this strategy.

When trying to add a more reliable pass I kept running into issues that I think will be easier to solve after we've made some of the discussed changes and generally fully moved over to the new IR.

[tybug]

lgtm! Shrinking benchmark is neutral:

Which isn't a bad thing, as this pass is likely to increase shrink quality, not just speed.

Instead of guessing where the one_of choices are, could we match the label of the node's example against OneOfStrategy? That should reduce false positives to zero.

[DRMacIver]

Instead of guessing where the one_of choices are, could we match the label of the node's example against OneOfStrategy? That should reduce false positives to zero.

So I thought about that, but I decided it was plausible that there were other patterns that had sufficiently similar shape that it was actually good to match things that aren't OneOfStrategy. e.g. you could implement the same sort of thing in a composite, but also this currently matches length parameters and that struck me as plausibly good for some hard to reduce cases of that.