The result of sob-hash function in the original reference implementation depends

on the order that hash-table-for-each visits entries, but I believe it should not.

The PR is a suggested fix--there may be a better one.

The change in sets-test.scm is the one exhibits the bug on Gauche, but whether

it reveals the bug or not depends on the implementation of hash tables.

I tested it with the change applied to Gauche source, but haven't tested

the PR itself.