cloudpickle drops __dict__ attribute of classes.

[pierreglaser]

After #253 (comment), I started examinating these lines. I think the original use-case was to handle classes like namedtuple overloading __dict__.
(even though the test suite passes for all python versions if simply drop the __dict__ item out of clsdict).

cloudpickle/cloudpickle/cloudpickle.py

Lines 633 to 635 in 167e163

	__dict__ = clsdict.pop('__dict__', None)
	if isinstance(__dict__, property):
	type_kwargs['__dict__'] = __dict__

In more detail:

• The first one line is necessary because updating a __dict__ field in the __dict__ of a class is simply not possible. Indeed, __dict__ attributes are proxies objects and do not support direct item assigment like __dict__[k] = v. Instead, the canonical way to update a class __dict__ is to do setattr(cls, key, value). However, here, key is __dict__, so python thinks we actually want to override the __dict__ attribute of cls, and not a __dict__ field of the __dict__.
• However, including __dict__ in the type_kwargs only if it is a property object is kind of arbitrary (probably related to namedtuples). Take this example for example, where we loose the __dict__ attribute:

In [1]: import gc 
 ...: import cloudpickle 
 ...: 
 ...: class A: 
 ...: __dict__ = {'some_attribute':1} 
 ...: a = A() 
 ...: print(f'calling __dict__ on the original A instance: {a.__dict__}') 
 ...: 
 ...: s = cloudpickle.dumps(a) 
 ...: del A 
 ...: del a 
 ...: gc.collect() 
 ...: depickled_a = cloudpickle.loads(s) 
 ...: print(f'calling __dict__ on the depickled A instance: {depickled_a.__dict__}')
calling __dict__ on the original A instance: {'some_attribute': 1}
calling __dict__ on the depickled A instance: {}

[ogrisel]

For reference, I tried with Python 2 and I observe the same outcome.