Support textless diarization in transcription details

[0xShug0]

Summary

Fixes #581.

The details endpoint inherits the transcription-only check and rejects pure diarization results before serializing their speaker turns. Relax that check only for diarization tasks on /v1/audio/transcriptions/details.

• Return an empty text value when diarization has no transcript, without dereferencing the absent optional.
• Accept valid silence results with no speaker turns; acceptance depends on the task rather than nonempty output arrays.
• Preserve the plain transcription endpoint, existing ASR responses, and the details endpoint streaming restriction.

The change is limited to app/server/runtime.cpp; no model or framework changes.

Validation

Rebuilt the Debug server and sent sequential real HTTP requests using CUDA and Sortformer v1 Q8 / Parakeet TDT Q8:

• Reproduced the Sortformer details failure before the change: HTTP 500, missing transcript text.
• After the change: speech returns HTTP 200 with speaker turns; silence returns HTTP 200 with empty text.
• JSON and multipart speech requests return identical speaker turns.
• Plain transcription still rejects textless diarization as before.
• stream=true on the details endpoint still returns HTTP 400.
• Parakeet plain and details responses match the baseline exactly after excluding timing fields.

Sortformer v2 was reviewed for the same result contract but was not runtime-tested.