feat(search): highlight the found line by meaning, without the query's words

[kdroidFilter]

A search now highlights the way the meaning finds it: in the found line alone, the passage closest in meaning, and none of the query's words. This applies to the results' snippets, their preview and the reader's smart find. A quoted search is unchanged: its words, everywhere.

Changes

• HybridSearchEngine
  • highlightsByMeaning is true for any non-quoted query.
  • Snippets are first shown without bold, then each hit's snippet is laid out on the passage of its found line, not of the neighbors that a short line's snippet carries.
• Preview (PassagePreview): neighbor lines are no longer highlighted; the found line shows the passage.
• Smart find: follows the same rule through SearchEngine.highlights (one logic for every view).
• seforim-semantic 0.5.0: finds that passage in ~6 ms per result instead of ~70. A static distillation of the model ranks the clauses; the model reranks only the best 2.
• SeforimLibrary: bumped to include the highlightsByMeaning hook and the hit line bounds (feat(search): highlight a found line by meaning, without the query's words SeforimLibrary#132).
• SearchBench: reports the refine cost per result and the process memory.

Measured (SearchBench, pages of 25, dev DB)

	value
highlight cost per result	~4 ms (goal: ≤ 10 ms)
native memory	+10 MB vs current master
static table (heap)	18.6 MB, by construction (not measured on the heap)
same passage as embedding every clause	89.7 %

Tests

• ktlint and detekt pass.
• jvmTest: everything passes except WidgetSizesScreenshotTest and CopyWithSourceE2ETest.talmud_multi_line_selection_trims_both_ends, which fail the same way on current master (unrelated to this change).
• Not checked visually in the running app.