The problem is that when committing a new file with empty contents (or
in general empty file with filelog p1 = -1), hg commit with narrow
doesn't create a filelog revision at all, which causes failures in
further commands.

The problem seems to be that:
- hg thinks that instead of creating a new filelog revision, it can
  use the filelog's p1 (the nullrev)
- because it thinks the file contents is the same in that revision and
  in p1
- because `narrowfilelog.cmp(nullrev, b'')` is True (unlike with
  `filelog.cmp`)

It's not clear to me which `cmp` behaves better. But I think it makes
sense to change the commit code to not to "reuse" the null rev when
adding an empty file with filelog p1 == filelog p2 == -1. This is
consistent with never writing the null rev in the manifest, which `hg
verify` claims is an invariant:

```
inside/c@4: manifest refers to unknown revision 000000000000
```

Differential Revision: https://phab.mercurial-scm.org/D11400