Evaluate impact of each filtering rule and drop the ones that are not needed

[michelole]

• At different ranks
• Different methods

[michelole]

For word2vec, the minimal set of rules that does not affect lenient metrics@10 is:

• acro not in full
• is_possible_expansion
• is_acronym_tail_on_last_word
• trim_plural (allows match NTX -> Nierentransplantation, but generates noise for BMS)

More rules might be needed if we're evaluating in a smaller rank (empirically, _is_relative_length_valid is also important to remove noise such as AINS: Ains)

not _has_capitals misses RV: rechtsventrikulär, but reduces overall noise.