discovery: compute newly discovered missing in a more efficient way

Calling "descendants" is expensive, instead, we bound the walk inside the know set
of undecided revision.

This help with discovery performance:

# without the revset '%ld' improvement
$ hg perfdiscovery -R pypy-left pypy-right
before: wall 0.675631 comb 0.680000 user 0.670000 sys 0.010000 (median of 15)
after:  wall 0.520145 comb 0.530000 user 0.510000 sys 0.020000 (median of 19)

There is another series in flight that greatly improves performances of "%ld"
substitution in `repo.revs` call. If this changeset is applied above it, we
see a similar performance boost.

# with the revset '%ld' improvement
$ hg perfdiscovery -R pypy-left pypy-right
before: wall 0.477848 comb 0.480000 user 0.480000 sys 0.000000 (median of 22)
after:  wall 0.404163 comb 0.400000 user 0.400000 sys 0.000000 (median of 24)

mercurial/setdiscovery.py

@@ -187,14 +187,10 b' class partialdiscovery(object):'
 
     def addmissings(self, missings):
         """registrer some nodes as missing"""
-        if self.missing:
-            new = self._repo.revs('descendants(%ld) - descendants(%ld)',
-                                  missings, self.missing)
-            self.missing.update(new)
-        else:
-            self.missing.update(self._repo.revs('descendants(%ld)', missings))
-
-        self.undecided.difference_update(self.missing)
+        newmissing = self._repo.revs('%ld::%ld', missings, self.undecided)
+        if newmissing:
+            self.missing.update(newmissing)
+            self.undecided.difference_update(newmissing)
 
     def addinfo(self, sample):
         """consume an iterable of (rev, known) tuples"""