Fix deadlocks in DecentralizedAverager and MPFuture

[borzunov]

This PR does the following:

1. Fix a bug in DecentralizedAverager.rpc_join_group().

   Previously, if a peer didn't manage to update its group key in the DHT in time (or didn't manage to propagate the updates), another peer could have chosen it as a leader based on the group key from the previous stage. The groups have been assembling incorrectly, which led to occasional deadlocks in tests (where the time periods between sequential averager steps are low).

   The fix forces leaders to compare followers' group keys before allowing them to join.

2. Fix a bug with MPFuture state after killing child processes.

   Many tests don't perform graceful shutdowns, so we are using a fixture that kills leftover child processes (if they don't respond to terminate() in 1 sec). Such killings may leave the global MPFuture state broken (e.g. the global locks acquired and the background thread blocked).

   The fix resets MPFuture backend (including its global state) after killing the child processes between tests.

3. Add -v flag to pytest in CI.

   Without this flag, we had no opportunity to determine a particular test that caused a deadlock since pytest used one output line per file (and Github CI erases the last output line when timeout was exceeded).

   The fix makes pytest to use one output line per test.

[codecov]

Codecov Report

Merging #331 (7c49501) into master (407c201) will decrease coverage by 0.02%.
The diff coverage is 77.77%.

@@            Coverage Diff             @@
##           master     #331      +/-   ##
==========================================
- Coverage   82.25%   82.22%   -0.03%     
==========================================
  Files          66       66              
  Lines        5979     5987       +8     
==========================================
+ Hits         4918     4923       +5     
- Misses       1061     1064       +3     

Impacted Files	Coverage Δ
hivemind/averaging/matchmaking.py	76.75% <50.00%> (-0.17%)	⬇️
hivemind/utils/mpfuture.py	89.66% <85.71%> (-0.17%)	⬇️
hivemind/dht/node.py	92.02% <0.00%> (-1.21%)	⬇️
hivemind/p2p/p2p_daemon.py	95.00% <0.00%> (+0.35%)	⬆️
hivemind/utils/timed_storage.py	96.93% <0.00%> (+1.02%)	⬆️
hivemind/averaging/key_manager.py	97.72% <0.00%> (+2.27%)	⬆️

[mryab]

Import from utils as well?

[borzunov]

In general, I'd weakly argue for avoiding importing from hivemind.utils directly to avoid having one module with a lot of mixed-purpose stuff and the corresponding long import line.

To follow this, I've changed from hivemind.utils import get_logger above to from hivemind.utils.logging import get_logger.