caption word chunking and text assembly duplicated across four places #3274

The caption word utilities (caption_text_from_words, the attach-to-previous helpers, normalize_caption_words, caption_word_chunks, and the segment size constants) are copy-pasted in four places and they've already drifted.

• crates/project/src/caption_timing.rs (private copy of the attach helpers + caption_text_from_words)
• apps/desktop/src-tauri/src/captions.rs (full set: attach helpers, normalize, chunk, text assembly, TARGET_CAPTION_WORDS_PER_SEGMENT etc.)
• apps/desktop-gpui/src/transcription.rs (same full set again)
• apps/desktop-gpui/src/editor_window.rs (another caption_text_from_words with the attach list inlined)

Concrete drift: the weak-word boundary check in captions.rs uses normalized.len() <= 1 (bytes) while transcription.rs uses normalized.chars().count() <= 1. So the unicode fix from #2685 only landed in the gpui backend, a one letter word like "à" still counts as two letters on the Tauri side.

I'd move the whole set into cap-project next to CaptionWord and have everything consume it from there. Happy to send that PR, and there's a second one I'd build on top of it (configurable words per caption, issue coming).

richie said on discord he's fine with PRs for this area.