Add simplemma analyzer

[osma]

Fixes #590

This PR adds a simplemma analyzer based on the Simplemma lemmatizer. Opening as a draft PR to get feedback from QA tools. The analyzer itself needs more testing, e.g. the quality of results and how it works with multiprocessing.

[codecov]

Codecov Report

Merging #591 (c8f94a7) into master (d07243a) will increase coverage by 0.00%.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##           master     #591   +/-   ##
=======================================
  Coverage   99.48%   99.48%           
=======================================
  Files          84       86    +2     
  Lines        5615     5645   +30     
=======================================
+ Hits         5586     5616   +30     
  Misses         29       29           

Impacted Files	Coverage Δ
annif/analyzer/__init__.py	93.93% <100.00%> (+0.39%)	⬆️
annif/analyzer/simplemma.py	100.00% <100.00%> (ø)
tests/test_analyzer_simplemma.py	100.00% <100.00%> (ø)

---

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 d07243a...c8f94a7. 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 test runs with MLLM and Parabel. There were no problems with parallellization. The initial version was a bit slow and memory-hungry so I implemented two optimizations, both copied from the Voikko analyzer:

1. use functools.lru_cache to cache lemmatize lookups, which is a bit faster (negligible memory cost)
2. avoid pickling the langdata object (happens with multiprocessing), instead load it from scratch using simplemma.load_data every time it is needed (e.g. in subprocesses) - this saves a lot of memory and is also faster

Table of test results:

backend	train set	lang	analyzer	train jobs	train time	train RSS	eval set	eval jobs	eval time	eval RSS	F1@5	NDCG
mllm	kirjastonhoitaja/fulltext-train	fi	voikko	4	357.78	1123776	kirjastonhoitaja/test	4	192.90	340248	0.3259	0.4368
mllm	kirjastonhoitaja/fulltext-train	fi	simplemma-initial	4	362.77	3456424	kirjastonhoitaja/test	4	221.09	2691512	0.3269	0.4303
mllm	kirjastonhoitaja/fulltext-train	fi	simplemma+lru_cache	4	357.15	3465288	kirjastonhoitaja/test	4	220.12	2691228	0.3217	0.4351
mllm	kirjastonhoitaja/fulltext-train	fi	simplemma+lru_cache+load_data	4	330.41	1900016	kirjastonhoitaja/test	4	213.71	1291668	0.3246	0.4363
omikuji-parabel	yso-finna (100k)	fi	voikko	4	243.78	748772	kirjastonhoitaja/test	4	21.84	557580	0.2033	0.2903
omikuji-parabel	yso-finna (100k)	fi	simplemma-initial	4	305.54	2748572	kirjastonhoitaja/test	4	49.17	2738524	0.2043	0.2890
omikuji-parabel	yso-finna (100k)	fi	simplemma+lru_cache	4	290.90	2775552	kirjastonhoitaja/test	4	41.80	2738080	0.2123	0.2826
omikuji-parabel	yso-finna (100k)	fi	simplemma+lru_cache+load_data	4	233.96	1648580	kirjastonhoitaja/test	4	42.36	1531068	0.2071	0.2970

Observations from the table:

• with the optimizations, simplemma is approximately as fast as voikko, but needs a bit more memory
• quality of results seems to be similar

I think this is good enough for review and subsequent merging unless any new problems surface.

[juhoinkinen]

LGTM

[osma]

I added documentation about the simplemma analyzer in the wiki.