Optimize CTE-heavy lineage by sharing scopes via Rc

[ashambalev]

I've run into performance issue in my project - parsing lineage in a complex dbt project with many CTEs was taking too much time (relatively, still faster than dbt itself). So checked lineage.rs and found the culprit:

to_node_inner deep-cloned every ancestor CTE scope (full expression trees) at each recursion level, and resolve_qualified_column re-cloned the entire set again for every column reference resolved through a CTE.

This PR switches the lineage recursion plumbing to Rc<Scope>.
Also self-exclusion filter now has a cheap pre-check by name, while still falling back to expression equality.

I've run a benchmark (on M5 Pro):

CTEs	main	PR	speedup
4	281 µs	135 µs	2.1×
8	1.78 ms	410 µs	4.4×
16	17.8 ms	1.63 ms	10.9×
24	67.1 ms	4.74 ms	14.1×

[tobilg]

Thanks a lot for the PR, merged! THis will be included in the upcoming version 0.6.2.

[tobilg]

I added some smallish improvements on top of it as GPT-5.6 xhigh indicated… Performance should be improved by 20-70% on a per-case basis.