Remove sync notifiers for a major speedup

[x42005e1f]

Avoids creating new threads since the mutex is already held in the queue methods.

[x42005e1f]

threading.Condition can be notified immediately without creating a new thread, since its notify() method is a very fast non-blocking operation, and the current implementation guarantees exclusive access to both _sync_not_empty and _sync_not_full. So synchronous notifiers were never really needed, and that is why they are removed in this PR. With this change, Janus became faster than the naive solution.

[codspeed-hq]

CodSpeed Performance Report

Merging #714 will not alter performance

_{Comparing x42005e1f:master (f058329) with master (e180116)}

Summary

✅ 4 untouched benchmarks

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 94.49%. Comparing base (e180116) to head (f058329).
Report is 1 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #714      +/-   ##
==========================================
- Coverage   94.54%   94.49%   -0.06%     
==========================================
  Files           5        5              
  Lines        1504     1490      -14     
  Branches      131      131              
==========================================
- Hits         1422     1408      -14     
  Misses         53       53              
  Partials       29       29              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[x42005e1f]

It looks like benchmarks are hardware dependent. In the CodSpeed profiler, I do not see notifiers for the test_bench_async_put_sync_get() case even before this PR, so I can assume that the thread has time to put items into the queue without blocking, and therefore does not call the notification methods. On my hardware, that test speeds up by ~5x.

[asvetlov]

It looks like benchmarks are hardware dependent.

codspeed does the best. It normalizes the speed to mitigate different performance of test boxes, and it also removes syscalls times.
The latest could be the reason of non-measurable changes in benchmarks.
There is walltime mode: https://docs.codspeed.io/instruments/walltime/ but I never tried it yet.

[asvetlov]

Good finding, thanks!