`map_rows(njit=True)` Improvements

[dougbrn]

This issue suggests a number of improvements to the map_rows njit feature implemented in: #451

• put comments in tests to the with pytest.raises blocks to explain why we expect it to fail.
• in tests use numpy.testing.assert_allclose/assert_array_equal and friends instead of for i in range(len(result1)): blocks
• Early exits with return would reduce amount of conditional blocks in _apply_njit_map_rows
• Leave _apply_njit_map_rows results as is without converting to lists, to avoid translation costs
• use .list_offsets for nestedseries directly

[Graciaaa3]

Already update comments to the pytest.raises blocks.

[hombit]

Please make numba an optional dependency.
It is actually a dev dependency