Calculate all pivot points based on end pivot

Previously, monthly pivots calculated from the end of a month failed to
account for the variation in the number of days per month. This was due
to subtle differences between subtracting an isodate.Duration against a
decremental date (previous implementation) and subtracting an
isodate.Duration with a multiplier against a fixed date (new
implementation).

Tests updated with the new behavior.

augur/frequency_estimators.py

@@ -79,7 +79,7 @@ def get_pivots(observations, pivot_interval, start_date=None, end_date=None, piv
     pivot = end
     while pivot >= start:
         pivots.appendleft(pivot)
-        pivot = pivot - delta
+        pivot = end - delta * len(pivots)
 
     pivots = np.array([numeric_date(pivot) for pivot in pivots])
 

tests/test_frequencies.py

@@ -150,17 +150,23 @@ def test_get_pivots_by_invalid_unit():
         (
             "2022-01-31",
             "2022-03-31",
-            ("2022-02-28", "2022-03-31")
+            ("2022-01-31", "2022-02-28", "2022-03-31")
         ),
+        # Note that Jan 31 to Apr 30 gives the same amount of pivot points as
+        # Jan 31 to Mar 31.
         (
             "2022-01-31",
             "2022-04-30",
             ("2022-02-28", "2022-03-30", "2022-04-30")
         ),
+        # However, in practice, the interval is more likely to be Jan 30 to Apr
+        # 30 as long as the start date is calculated relative to the end date
+        # (i.e. start date = 3 months before Apr 30 = Jan 30).
+        # That interval includes an additional pivot point as expected.
         (
             "2022-01-30",
             "2022-04-30",
-            ("2022-02-28", "2022-03-30", "2022-04-30")
+            ("2022-01-30", "2022-02-28", "2022-03-30", "2022-04-30")
         ),
     ]
 )