sor path finding optimisations

Why this is faster

- No recursion or repeated array copies in token-path search; we push/pop on a single shared path and visited set.
- Incremental constraint tracking avoids O(len(path)) scans on each step.
- Early stop based on approxPathsToReturn (as your comments suggest) keeps the search bounded.
- Beam search on ranks uses cached per-edge limits to prune aggressively before running the expensive full-path limit calculation.
- At most approxPathsToReturn full getLimitAmountSwapForPath evaluations per token path, instead of exploring all rank combinations.

Author: gmbronco