Make the imputation iterative

[nicolassidoux]

• When given a dataset to impute, the imputer will check if a big enough index can be built. If not, it will temporarily exclude enough features to reach this threshold and try again. If an index still can't be built, it will use fallbacks values as defined by the user strategy.
• The "big enough index" is decided by threshold that can be passed when constructing the imputer.
• The imputation will be performed again for any feature that was left behind in the previous step, making the process iterative.
• Use fallback strategy for extremely sparse data where data is simply (= mean/median) imputed (as implemented by @Zethson) if an index can't be built at all.
• The tests were improved to include some common tests as used in ehrapy, also @Zethson added a lot of tests for edge cases.

[Zethson]

Generally great! Let's see what your benchmarks give us.

It's midnight and I'm too tired to verify the correctness of your implementation but I can have another look after the benchmarks.

[codecov]

Codecov Report

Attention: Patch coverage is 91.01124% with 8 lines in your changes missing coverage. Please review.

Please upload report for BASE (main@8fb7fe6). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
src/fknni/faiss/faiss.py	91.01%	8 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##             main       #8   +/-   ##
=======================================
  Coverage        ?   92.45%           
=======================================
  Files           ?        3           
  Lines           ?      106           
  Branches        ?        0           
=======================================
  Hits            ?       98           
  Misses          ?        8           
  Partials        ?        0           

Files with missing lines	Coverage Δ
src/fknni/faiss/faiss.py	92.07% <91.01%> (ø)

[Zethson]

Just a few more minor things and I'd love to merge this. Thank you!