Improve handling of compound words in search

[noeddl]

Fixes #821
Depends on okfde/froide#1122 (both PRs should be reviewed together)

This PR improves the handling of compound words and reduces false positives in search.

Changes

• Compound word handling

  • Preprocesses search queries so that subtokens of compound words are connected with AND instead of OR (as described in this article). This ensures that all subtokens must be present in a document for it to match.
  • Uses the output of the decompounder token filter in search queries only if the subtokens actually form the full compound word.
  • Prevents false positives caused by very short subtokens and nested subtokens.

• General search improvements

  • Removes stop words from non-exact queries.
  • Ensures quoted phrases only match exact results.
  • Adds tests to cover and document the current search behavior.

Notes

• Search is still far from perfect. The decompounder algorithm itself could be improved, but this likely requires significant effort (e.g. a custom Elasticsearch/Lucene plugin).
• A lighter alternative for improvement could be extending the decompounder dictionary.
• Additional improvements could be achieved by using a more sophisticated stemmer or lemmatizer.
• Longer-term, we may want to consider switching to more advanced approaches such as semantic search.