Whisper word timestamps wrong for no-space languages (potential fix included)

[hlevring]

Add_word_level_timestamps() always groups subword tokens with
split_tokens_on_spaces(), which depends on leading spaces to find word boundaries

This does not work for no-space (Scriptio continua) languages (Chinese, Japanese, Thai, Lao, Burmese, Cantonese) where decoded text has no spaces.

I made a fix in 2026.2 branch of my own fork which seems to work (afaik)
cea5a80

I could potentially make a pull request against main, but the changes are pretty small and only in a single file, so it might be easier for maintainers to just fix directly.

[as-suvorov]

@hlevring I'll prepare a PR as soon as I'll be available after code freeze. Thanks a lot!

[as-suvorov]

@hlevring Could you please double check implementation in #4131 works for you?

[hlevring]

@as-suvorov Yeah, looks good. And nicer than my quick fix for lang_to_id lookup ;) Thanks