Perform suggest operations in parallel using multiprocessing in nn_ensemble

[osma]

During training, the NN ensemble has to process the training documents through the source projects, collecting suggestions. Previously this was done sequentially and could take a lot of time, often dominating the time it takes to train a NN ensemble.

With this PR, the suggest operations are done in parallel on multiple CPUs, controlled by the --jobs parameter of the annif train command. The default is to use all CPUs.

Fixes #429

Opening a draft PR to get feedback from QA tools. More detailed testing and benchmarking still needs to be done.

[codecov]

Codecov Report

Merging #568 (f2cf2ea) into master (736834b) will increase coverage by 0.00%.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##           master     #568   +/-   ##
=======================================
  Coverage   99.47%   99.47%           
=======================================
  Files          84       84           
  Lines        5556     5565    +9     
=======================================
+ Hits         5527     5536    +9     
  Misses         29       29           

Impacted Files	Coverage Δ
annif/backend/nn_ensemble.py	99.30% <100.00%> (+0.04%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 736834b...f2cf2ea. Read the comment docs.

[sonarcloud]

Kudos, SonarCloud Quality Gate passed!   

0 Bugs
0 Vulnerabilities
0 Security Hotspots
0 Code Smells

No Coverage information
0.0% Duplication

[osma]

I did some benchmarks using the yso-nlf data set from the Annif tutorial. I set up TFIDF, MLLM and Omikuji projects according to the tutorial exercises, then created a NN ensemble project including all three as sources, like this:

sources=yso-tfidf-en,yso-mllm-en:2,yso-omikuji-parabel-en

I trained the NN ensemble using 500 documents (--docs-limit 500) with different setups on an i7-8550U laptop (4 cores, 8 threads). In the end I also evaluated the ensemble project against the test set, to verify that the ensemble is still working. Results:

	wall time	user time	max RSS	F1@5
Before PR	22:46	1448	3021780	0.3632
This PR, jobs=8	10:31	4846	2643096	0.3666
This PR, jobs=1	22:31	1442	3312252	0.3727

Observations:

• Training time decreases significantly (to less than half) when using parallel processing, although there is significant overhead in the parallelization - the total CPU time (user time) increases a lot.
• With jobs=1, the situation is similar to what it was before.
• Memory usage does not change dramatically.
• Evaluation scores are similar to what they were before.

[juhoinkinen]

Great improvement!