DEV-1890 whisper words - #19

…sults

The transcription form only had include_ts, so the v2 parameters added
alongside whisper-ct2 were unreachable from the tester, and a structured
result was just an opaque download link.

Form (all four /audio/transcriptions endpoints, via a shared builder):
- ts_level, diarize, lang, and include_metadata (URL sources only -- the
  API returns null metadata for an uploaded file)
- capability-gated rather than hardcoded: ts_level appears only for models
  publishing info.limits.timestamp_levels and takes its options from that
  list, diarize only for info.features.supports_diarization. Two new
  registry hooks carry this: optionsFromModel and visibleFromModel.
- ts_level is clamped to the selected model's list, so switching models
  can't leave a value the API rejects
- visibleWhen now accepts an array of conditions (OR) and compares on the
  string form, so it works for booleans: ts_level shows for include_ts OR
  diarize. Hidden fields stay out of the payload, price calc included.

Results: the shape is the model's choice, reported as `structured` on the
job status. The transcript strip shows JSON/TXT, language, delivered
ts_level and whether speakers came back; expanding renders plain text as
text, and a structured result as segments with ms times, coloured
speakers, avg_logprob and per-word spans. Source metadata renders as a
key/value block, generically, so new fields need no code change.

Result files are fetched through a new /api/result -- result URLs are
presigned and unreachable from the browser -- and it names the host on a
connection failure, since an internal storage hostname that only resolves
inside the API's network is otherwise a bare "fetch failed".

Layout: the toggles move from the right column to the centre one.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>