[LFSC] Fix performance issues, more determinism

[LFSC] Fix performance issues, more determinism

For certain proofs, the performance was drastically different on
different OSes. The cause for this difference was a pointer comparison
in the deduplication in `Expr::defeq()`. Depending on how the OS
allocated the memory, an expression `a` would get replaced with an
equivalent expression `b` or vice versa, which in turn affected
performance of `Expr::free_in()` dramatically (sub-second vs. >5 min
running times). This commit makes the process more deterministic by
using a heuristic that favors symbolic expressions and greedily tries to
make small refcounts smaller when replacing, and changes
`Expr::free_in()` to not repeatedly explore the same subexpressions.