Preserve spacing of toplevel comments

[gpetiot]

Motivation:

let _ = assert (foo (bar + baz <= quux))
(* this comment should stay attached to the preceding item *)

let _ = foo

was formatted as:

let _ = assert (foo (bar + baz <= quux))

(* this comment should stay attached to the preceding item *)

let _ = foo

because the comment was wrongfully attached to the second item instead of the first one.

edit: no diff with test_branch.sh after the unnecessary commits were dropped.

[gpetiot]

Waiting for #1549

[Julow]

Looks good.

[Julow]

This function doesn't make sense if fst is after (both start and end positions) snd. I think we should at least test assert (fst.loc_start.pos_lnum <= snd.loc_start.pos_lnum).
Also I'm not sure what this function means if snd is inside fst, we should probably assert that the result is >= 0 too.
Could you add some documentation about that ?

[gpetiot]

I agree on the principle, but I've just discovered that this invariant doesn't hold anymore when comments have been relocated, so for now I've just added a docstring for this function and I will investigate more deeply this relocation issue in another PR. It doesn't look blocking for this PR as this is what was already done in master (the difference computation) anyway.