Nondeterminism in the serailizer/deserializer

[FTRobbin]

There is some nondeterminism in serializing/deserializing the egraph that requires sorting the tables to normalize. Without it, the idempotence test fails. I think nondeterminism should be avoided at the root if possible.

[FTRobbin]

Related: egraphs-good#793

So the issue might be in egglog.

[FTRobbin]

Observed more of this issue when running the extract experiment on Herbie workflows.

The problem is intrinsic to egglog.