cleanup: fix leakage of dirstate._map to client code
We already had proper accessors for most of the behavior of
dirstate._map that callers cared about exposed in the actual dirstate
class as public methods. Sigh.
There are two remaining privacy violations in the codebase after this change:
1) In the perf extension, which I suspect has to stick around because
it's really testing the dirstate implementation directly
2) In largefiles, where we deal with standins and mutating status. Looking at
this, I _strongly_ suspect a formal dirstate interface would allow
this to work via composition instead of inheritance and
monkeypatching. Fortunately, such wins are a part of my motivation
for this work. I anticipate we'll come back to this in due time.
Differential Revision: https://phab.mercurial-scm.org/D6837