Fix overly-aggressive deduplicator

[pavpanchekha]

This PR fixes a pretty major bug in Herbie, noted in #1372: Herbie has recently turned terrible at rewrites. I figured out why; it's the deduplicator. Basically, suppose we have a batch with both the expressions e1 = 1 and e2 = (1 + x) - x in it. (In fact e1 is a subexpression of e2.) Ok, well, we will prove that e2 = e1 in the rewriter. So we will construct the alt histories e1 -> e1 and also e2 -> e1. However, unfortunately, we then deduplicate by final expression, deleting the second (useful) alt and keeping only the first (dumb) one.

This PR fixes it in a fairly ham-fisted way by deduplicating less, which might mean Herbie runs for longer. But it's surely worth it.

[pavpanchekha]

Fun fact: the reason for the weird branch name is that I was convinced the issue was something about how the egraph is growing too fast, and was ready to do a super deep dive. And then it turns out... it's this other thing!

[pavpanchekha]

Nightly results (vs main) are good but not life-changing. I still think we should merge, at least to fix #1372 but more importantly because I think this but is a big part of why Herbie does poorly without series expansion.

[AYadrov]

I think I got it - we actually want to prune a substitution of a final in initial alt, it is too early to make any decision right after rewriting stage - we actually want to prune after reconstruct!, or, in other words, we want to prune after rewriting stage by (cons initial final) key