Support full featured groupby function

Currently df.groupby() in Mars only implements ``as_index``. ``sort`` is added but not implemented yet, while ``level`` is not implemented which is also useful. What's more, when ``GroupBy`` object is generated, keys and grouped data frames are serialized instead of ``GroupBy`` object itself, which may not be a universal solution for different scenarios of groupby.

``sort_index`` (#1037) may be useful when implementing this.

Subtasks:

* [x] Fix serialization issue of GroupBy objects (#1136)
* [x] Support groupby-getitem (#1136)
* [x] Support level option (#1136)
* [x] Support ``group_keys`` (#1136)
* [ ] Support ``observed``
* [x] Support ``sort`` (#2959)
* [x] Support args of cumulative functions in groupby (and integrate ``df.sum(level=x, **kwargs)``)
* [x] Support grouping by series (#1181)

Start at Mars's df.groupby() implementation and compare the remaining observed option with the existing groupby options. Done means observed is supported for groupby; the issue names no specific file or test path, so those will need to be located first.

Written by the indexing model from the issue text.