Implement custom MeanLayer in nn_ensemble

[osma]

This PR replaces the Lambda layer with a custom MeanLayer in the nn_ensemble backend. It was intended as a fix to the CustomMaskWarning error encountered in PR #499 (upgrade to TensorFlow 2.5.0) but it turned out that the problem is unrelated.

This might still be a good idea, for example it could improve the compatibility of saved nn_ensemble models between Python versions. But it needs more testing so I'm leaving this as a draft PR.

[codecov]

Codecov Report

Merging #500 (2403423) into master (0fe6557) will increase coverage by 0.00%.
The diff coverage is 83.33%.

@@           Coverage Diff           @@
##           master     #500   +/-   ##
=======================================
  Coverage   99.48%   99.48%           
=======================================
  Files          78       78           
  Lines        5669     5672    +3     
=======================================
+ Hits         5640     5643    +3     
  Misses         29       29           

Impacted Files	Coverage Δ
annif/backend/nn_ensemble.py	99.23% <83.33%> (-0.77%)	⬇️
annif/backend/stwfsa.py	100.00% <0.00%> (+1.51%)	⬆️

---

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 0fe6557...2403423. 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 tested this in a situation where Python 3.6 is used for training models and Python 3.8 for using them. The PR seems to improve compatibility between Python versions for NN ensemble projects, but both instances need to run the updated code.

First I shamelessly copied the minimal setup from #453:

# projects.cfg:
[arch-fasttext]
name=FastText
language=fi
backend=fasttext
limit=100
vocab=arch
analyzer=snowball(finnish)

[arch-nn]
name=YSO NN ensemble Finnish
language=fi
backend=nn_ensemble
sources=arch-fasttext
limit=100
vocab=arch
nodes=100
dropout_rate=0.2
epochs=2

# Training on Python 3.6:
annif loadvoc arch-fasttext tests/corpora/archaeology/subjects.tsv
annif train arch-fasttext tests/corpora/archaeology/documents.tsv
annif train arch-nn tests/corpora/archaeology/fulltext/

# Testing on Python 3.8:
echo arkeologia | annif suggest arch-nn

Without this PR, the annif suggest command fails with this error:

ValueError: bad marshal data (unknown type code)

With this PR applied, it works without problems.

I also tested the case of old models trained without this PR. They still keep working with the PR applied. So this PR doesn't break compatibility for old NN ensemble projects, but projects will have to be retrained to be able to benefit from the additional compatibility this PR offers.

[juhoinkinen]

Like above I tested training models on Python 3.7 and using them on 3.9: without this PR the bad marshal data (unknown type code) error appears, but with this PR applied it does not (suggest call works).