Fix minor asyncio issues in averager

[borzunov]

This fixes minor asyncio issues in averager found while debugging #351:

1. Make averager not fail during reporting an error if the recipient closes the connection before receiving it (e.g. if it has crashed itself). Example:

   

2. Support asyncio cancellation and shutdown in _load_state_from_peers, that is don't suppress asyncio.CancelledError and GeneratorExit. For example, this happens when GeneratorExit is suppressed:

   

3. Remove excess calls of __aiter__()/__anext__() (outside hivemind.utils.asyncio, we can always manage with anext(), so let's not scare the code readers with these magic methods)

[codecov]

Codecov Report

Merging #356 (3789b7e) into master (4da80bc) will increase coverage by 0.18%.
The diff coverage is 83.33%.

@@            Coverage Diff             @@
##           master     #356      +/-   ##
==========================================
+ Coverage   83.19%   83.38%   +0.18%     
==========================================
  Files          66       66              
  Lines        5994     6002       +8     
==========================================
+ Hits         4987     5005      +18     
+ Misses       1007      997      -10     

Impacted Files	Coverage Δ
hivemind/averaging/allreduce.py	77.70% <0.00%> (ø)
hivemind/averaging/averager.py	84.73% <0.00%> (ø)
hivemind/averaging/matchmaking.py	84.98% <ø> (+1.20%)	⬆️
hivemind/p2p/p2p_daemon.py	97.01% <100.00%> (+0.04%)	⬆️
hivemind/utils/asyncio.py	95.89% <100.00%> (+1.68%)	⬆️
hivemind/dht/node.py	92.60% <0.00%> (+1.19%)	⬆️