Implement CenteredClip in averager

[borzunov]

• (important) Reducers should work concurrently
• (important) Test with .isfinite(x), test that no other tensor values may corrupt CClip
• Rename to MeanReducer, CenteredClipReducer (btw, why not *Aggregator?)
• run_in_executor for reducers
• nit: Rename Factory to smth else, make_reducer: Callable[...]
• n_steps -> max_steps

[codecov]

Codecov Report

Merging #379 (07d8f5f) into master (b84f62b) will decrease coverage by 0.30%.
The diff coverage is 53.62%.

❗ Current head 07d8f5f differs from pull request most recent head e9f3288. Consider uploading reports for the commit e9f3288 to get more accurate results

@@            Coverage Diff             @@
##           master     #379      +/-   ##
==========================================
- Coverage   83.98%   83.68%   -0.31%     
==========================================
  Files          70       71       +1     
  Lines        6383     6443      +60     
==========================================
+ Hits         5361     5392      +31     
- Misses       1022     1051      +29     

Impacted Files	Coverage Δ
hivemind/averaging/accumulators.py	45.76% <45.76%> (ø)
hivemind/averaging/allreduce.py	77.56% <100.00%> (+0.14%)	⬆️
hivemind/averaging/averager.py	86.27% <100.00%> (+0.03%)	⬆️
hivemind/averaging/partition.py	98.03% <100.00%> (-0.02%)	⬇️
hivemind/averaging/matchmaking.py	83.75% <0.00%> (-0.63%)	⬇️
hivemind/dht/node.py	92.63% <0.00%> (+1.18%)	⬆️

[justheuristic]

preference: let's default to some reasonable delta and a very large n steps

[justheuristic]

Random thoughts:

• gotta ensure that we do not fail even if user sends nan/inf/near_inf values (both with compression and with raw values)
• would prefer to set delta to 1e-6 and increase n_iters dramatically (e.g. to 50) to make runtime adaptive
• let's print some warnings in case your values were more than e.g. 3 tau outside the range

[borzunov]

Suggested change

	diff.addmm_(-coeffs.repeat(n_peers, 1), diff)
	diff -= coeffs @ diff

It seems like addmm_() doesn't work correctly if the destination is equal to one of the operands.