Improved handling of source weights in NN ensemble

[osma]

This PR contains fixes aiming at improving the way the NN ensemble handles sources that have been assigned non-default weights.

1. The weights assigned to source backends are now always normalized so that they add up to 1 (L1 normalization).
2. The NN ensemble multiplies incoming score vectors by the assigned weights. Since the normalized weights are now generally smaller than they used to be (e.g. not 1:1:1 but 0.33:0.33:0.33), they are multiplied - outside the neural network - by the number of sources to compensate for this change.
3. While researching this, I discovered that the NN ensemble performs better when the incoming weights are slightly larger numerically. Taking the square root of the scores (which increases the scores, since sqrt(x) > x when 0<x<1) before feeding them to the neural network seems to improve F1@5 scores by an additional 1-2 percentage points!

Fixes #457

[sonarcloud]

Kudos, SonarCloud Quality Gate passed!

0 Bugs
0 Vulnerabilities
0 Security Hotspots
0 Code Smells

No Coverage information
0.0% Duplication

[codecov]

Codecov Report

Merging #458 (90f3f76) into master (98bde23) will increase coverage by 0.00%.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##           master     #458   +/-   ##
=======================================
  Coverage   99.41%   99.41%           
=======================================
  Files          65       65           
  Lines        4627     4630    +3     
=======================================
+ Hits         4600     4603    +3     
  Misses         27       27           

Impacted Files	Coverage Δ
annif/backend/nn_ensemble.py	100.00% <100.00%> (ø)
annif/util.py	97.72% <100.00%> (+0.10%)	⬆️

---

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 98bde23...90f3f76. Read the comment docs.

[osma]

There's also the question what happens if you use an old model, trained with the current/old NN ensemble code, while running the new code. I tested that with the example model from #457 that originally achieved an average F1@5 score of 0.4418; when evaluated with the new code, the score was 0.4330, which is a bit worse, but not catastrophically so - and in fact for some of the individual data sets the scores had improved.

I think it's enough to state in the release notes that retraining NN ensemble models is recommended, but not absolutely necessary.

[osma]

I verified that this PR doesn't change the results of a simple ensemble. (The annif.util.parse_sources that was modified in this PR is used by all ensembles)

Still need to double-check that PAV isn't affected

[osma]

I verified this once more with the Annif tutorial data sets. The results improved significantly for the STW data set, but declined slightly for the YSO one (though depends on metric).

Double-checked that PAV is unaffected. Merging this now.