refactor: sort/dedup lineage nodes once, not per-node

[prql-bot]

Problem

FrameCollector::fold_expr (in semantic/reporting.rs) is invoked once per expression node while building the debug-lineage node graph. On every one of those calls it re-ran:

self.nodes.sort_by_key(|a| a.id);
self.nodes.dedup();

Since fold_expr runs n times and each call sorts the growing nodes vector, collection was O(n² log n) for a query with n nodes — pure wasted work that scales badly on large queries. Nothing reads self.nodes mid-traversal, so the intermediate sorts served no purpose; only the final ordered/deduped vector is used (by collect_frames, for the parent-linking pass).

Solution

Move the sort + dedup out of the per-node fold_expr and into collect_frames, running them once after the full traversal completes (mirroring how collect_frames already does frames.reverse() and parent-linking after folding). This drops the work to a single O(n log n) sort. The final nodes vector is byte-for-byte identical, because a single sort+dedup after all pushes produces the same canonical result the repeated in-loop sorts converged to.

Testing

This is a behavior-preserving change, so there is no "fails-before / passes-after" regression test to add — output is unchanged by construction. Correctness is guarded by the existing debug_lineage integration snapshots, which serialize the full node graph that collect_frames produces:

• All 31 queries::debug_lineage::* snapshot tests pass unchanged.
• The reporting/lineage lib tests pass.

Found during the nightly code-quality survey.