fix(tokenizer): preserve known-word highlight when POS filters suppress

- Known-word cache matches now set isKnown=true even for tokens excluded by POS filters
- POS exclusion gate suppresses N+1, frequency, and JLPT only; known status is computed before the gate
- Jellyfin subtitle preload continues after cleanup failures instead of aborting
- Update config docs and option description to document the known-word bypass behavior

config.example.jsonc

@@ -523,7 +523,7 @@
       "highlightEnabled": false, // Enable fast local highlighting for words already known in Anki. Values: true | false
       "refreshMinutes": 1440, // Minutes between known-word cache refreshes.
       "addMinedWordsImmediately": true, // Immediately append newly mined card words into the known-word cache. Values: true | false
-      "matchMode": "headword", // Known-word matching strategy for subtitle annotations. Values: headword | surface
+      "matchMode": "headword", // Known-word matching strategy for subtitle annotations. Cache matches always receive known-word highlighting even when POS filters suppress other annotation types. Values: headword | surface
       "decks": {} // Decks and fields for known-word cache. Object mapping deck names to arrays of field names to extract, e.g. { "Kaishi 1.5k": ["Word", "Word Reading"] }.
     }, // Known words setting.
     "behavior": {