`Object_graph.iter` needs cleaning

[Ngoguey42]

Aside from needing more docstrings, more tests (#1134) and maybe some benchmarks, the specifications of Object_graph.iter need to be improved.

Issues:

• Mark could be called more: https://github.com/mirage/irmin/pull/1220#discussion_r550484868
• The depth mechanism could be changed: https://github.com/mirage/irmin/pull/1220#discussion_r550484905
• Using iter_p might be problematic: https://github.com/mirage/irmin/pull/1220#discussion_r550484949
• The closure mechanism for branch doesn't work: https://github.com/mirage/irmin/pull/1220#discussion_r550484975
• The rev=false order of iteration is wrong: https://github.com/mirage/irmin/pull/1220#discussion_r550485036

Summaty of the 4 APIs of iter

Object_graph.S.iter
    cache_size ->
    depth ->
    pred ->
    node -> edge ->
    skip ->
    t -> min -> max -> rev -> unit -> unit Lwt.t

Node.Graph.iter
    node -> contents -> edge ->
    skip_node -> skip_contents ->
    t -> min -> max -> ?(rev = true) -> unit -> unit Lwt.t

Repo.iter
    cache_size ->
    edge branch commit node contents
    skip_branch -> skip_commit -> skip_node -> skip_contents ->
    pred_branch -> pred_commit -> pred_node -> pred_contents ->
    t -> min -> max -> ?(rev = true) -> unit Lwt.t

Commit.History.iter
    commit -> edge ->
    skip ->
    t -> min -> max -> ?(rev = false) -> unit -> unit Lwt.t

[samoht]

This needs a new scheduling label.

[Ngoguey42]

As discussed offline with @n-osborne a few months ago, iter would be more predictable if min was made simpler.

min should be the list of objects that iter will stop at (i.e. no special case).

A new boolean flag should control whether the mins are included or excluded from the iteration (i.e. whether or not they are passed to edge and node)