Let a group of nothing but nans take the fill value

[ml31415]

The nan forms of the pure python implementation filter the nan values out of the input before grouping. When every value of the input is a nan, nothing is left to filter down to and unpacking the empty remainder raises

ValueError: not enough values to unpack (expected 2, got 0)

A group that loses all of its values takes the fill_value - the same a group missing from group_idx does - and that is what the numpy implementations already make of such a group (0 for nansum, 1 for nanprod, 0 for nanlen, False for nanall, -1 for nanargmin, nan for the rest). Now the same here.

The new test compares that group against the numpy implementation over all fifteen nan forms, for three inputs: the whole input a nan, a single all-nan group next to one that has a value left, and the same with complex values - the filter compares val == val since #98 landed, so a nan in either part has to empty the group as well. Without the change it fails with the ValueError above; the mixed input alone does not trigger it, which is why the first and the third case are in there.

Two notes:

• The test calls aggregate_purepy directly. The parametrized tests do not reach the pure python nan forms: _is_implemented() reports them as not implemented, and that is more than a stale label - removing it makes 27 tests fail (12 test_var_with_nan_fill_value, 7 test_cmp, 2 test_default_fill_value, 5 complex). Untangling pure python's nan semantics is a project of its own and not part of this.
• Rebased onto Support complex input in the aggregation functions #98. It rewrote this very line to filter with val == val instead of math.isnan, which is what makes the complex case above possible; the two changes resolved into one block.

Measured: 2395 passed, 174 skipped, 924 deselected - one more than master, the new test.