Optimize min_count when expected_groups is not provided. (#236)

* Optimize min_count for all numpy

For pure numpy arrays, min_count=1 (xarray default) is the same
as min_count=None, with the right fill_value. This avoids
one useless pass over the data, and one useless copy.

We need to always accumulate count with dask, to make sure we
get the right values at the end.

* Better?

Author: dcherian

flox/core.py

@@ -2324,6 +2324,7 @@ def groupby_reduce(
     nby = len(bys)
     by_is_dask = tuple(is_duck_dask_array(b) for b in bys)
     any_by_dask = any(by_is_dask)
+    provided_expected = expected_groups is not None
 
     if (
         engine == "numbagg"
@@ -2440,7 +2441,7 @@ def groupby_reduce(
     #     The only way to do this consistently is mask out using min_count
     #     Consider np.sum([np.nan]) = np.nan, np.nansum([np.nan]) = 0
     if min_count is None:
-        if nax < by_.ndim or fill_value is not None:
+        if nax < by_.ndim or (fill_value is not None and provided_expected):
             min_count_: int = 1
         else:
             min_count_ = 0