perf(topological_sort): traverse iteratively and drop the redundant sets

[tachsin]

Part of #795, which covers the same defect in strongly_connected_components as well; that half is #799, and the issue can be closed once both have landed.

visit recursed once per node, so a deep graph overflowed the stack: a chain of 200 000 nodes is enough, and so is a random DAG of 60 000. This replaces the recursion with an explicit stack.

The bookkeeping around it was doing the same work more than once:

• marked and temp are nested — a node is in temp from the moment it is reached and in marked once it is finished, so marked is a subset of temp and every node was hashed into both. One map from node to "finished?" answers both questions in a single lookup.
• The outer loop repeatedly took an arbitrary key out of a HashSet of roots that visit was emptying, so each call rescanned the buckets left behind by the ones before it. Walking the caller's slice needs no set at all, and with it goes a clone of every root.
• Both sets used the default hasher, unlike the rest of the crate.

topological_sort_into_groups keeps its structure and only changes hasher.

Checking

successors is still called exactly once per node, which the existing complexity test covers. Orders are checked edge by edge against 200 random DAGs, and a regression test runs a 200 000 node chain on a deliberately small 1 MiB stack.

On 60 000 nodes: about 76% off topological_sort, about 33% off topological_sort_into_groups.

All timings measured on Windows 11, Intel Core Ultra 7 265K, 64 GB RAM.