Key observation: 99.9% of the time Futures and Tasks have only one callback.  

Currently we have a list of callbacks in each Future/Task.  We can avoid list object allocation if we add a field to store the first callback.  This way we'll only need to ever allocate the list if a Future/Task has more than one done callbacks.

asyncio with the proposed patch applied shows that 3-6% better performance of an asyncio echo server implemented using asyncio.streams.  This benchmark involves quite a bit of other Python code, so this improvement is actually quite significant.

The patch consists of:

1. first callback / callbacks list refactoring.

2. a free list implementation for Future.__await__ objects (which are created on almost all awaits in asyncio).

3. a big cleanup of the code, ensuring that Future/Task are always initialized properly.

New changeset 1b7c11ff0ee3efafbf5b38c3c6f37de5d63efb81 by Yury Selivanov in branch 'master':
bpo-32348: Optimize asyncio.Future schedule/add/remove callback. (#4907)
https://github.com/python/cpython/commit/1b7c11ff0ee3efafbf5b38c3c6f37de5d63efb81

The attached benchmark from https://bugs.python.org/issue32204 shows that the updated Task is 15% faster.