### Summary
Slovak and Czech datasets are effectively BM25-blind today. The tokenizer only keeps ASCII letter runs whole, so every accented word is fragmented before indexing (`škola` -> `š kola`) and the keyword leg of hybrid search never matches. Neither language has a Snowball stemmer either, so the analyzer inherits whatever stemmer the previously processed dataset left behind.
Co-authored-by: mcivan <martin.civan@nases.gov.sk>
### What problem does this PR solve?
Repeated stemmer initialization created new Snowball environments
without releasing the previous ones, causing native memory usage to grow
continuously during retrieval.
This change reuses the existing environment when the language is
unchanged and releases the previous environment before switching
languages.
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
Martin Civáňandmcivan
Anupam Mediratta
qinling0210
buua436
