rebase -r: preserve skipped ancestor edges

[bnjmnt4n]

Just commenting why I think the f->e->b edge should be preserved:

1. disconnected -r targets are rebased as if they were all connected
2. in order to achieve that, internal edges across skipped nodes are preserved
3. so it seems better to apply the same rule to external edges

In practice, I wouldn't try this kind of weird rebase, and I wouldn't mind if f->e->b was removed. My point is that it's easier to reason about the behavior if we have a consistent model (such as rebase -r c|f is equivalent to rebase --keep -r c::f + abandon c|f + abandon e'.)

    create_commit(&test_env, &repo_path, "a", &[]);
    create_commit(&test_env, &repo_path, "b", &["a"]);
    create_commit(&test_env, &repo_path, "c", &["a"]);
    create_commit(&test_env, &repo_path, "d", &["a"]);
    create_commit(&test_env, &repo_path, "e", &["c", "b"]);
    create_commit(&test_env, &repo_path, "f", &["e"]);
    insta::assert_snapshot!(get_log_output(&test_env, &repo_path), @r###"
    @  f
    ◉    e
    ├─╮
    │ ◉  b
    ◉ │  c
    ├─╯
    │ ◉  d
    ├─╯
    ◉  a
    ◉
    "###);

    // TODO: edge f->e->b is lost
    test_env.jj_cmd_ok(&repo_path, &["rebase", "-r", "c|f", "-d", "d"]);
    insta::assert_snapshot!(get_log_output(&test_env, &repo_path), @r###"
    @  f
    ◉  c
    ◉  d
    │ ◉  e
    ╭─┤
    │ ◉  b
    ├─╯
    ◉  a
    ◉
    "###);

Originally posted by @yuja in #3605 (comment)