v0.17.0

Oriel 0.17.0

Oriel tells voices apart: oriel.speakers, speaker embeddings and
offline diarization, the module behind GhostPen's meetings (who said
what, and named voices remembered across meetings).

Added

• oriel.speakers (.speakers = true, needs .llama or .whisper;
  Linux, Windows, macOS): speaker embeddings and offline diarization on
  voice-detect.cpp (MIT), compiled against the shared ggml build.
  Encoder loads a GGUF encoder from catalog (WeSpeaker ResNet34 by
  default, ECAPA-TDNN) and embeds 16 kHz speech; diarize embeds 3 s
  windows every 1.5 s over the speech regions, clusters them
  (average-linkage, cosine; or a known speaker count) and returns speaker
  turns with one centroid per speaker; cluster does the clustering on
  embeddings an app made itself; match finds a known voice for a
  centroid. Progress and cancellation for long recordings. On LibriSpeech
  voices WeSpeaker put one man's halves at cosine 0.73, two women at 0.22,
  a man and a woman at 0.08 (the ECAPA GGUF: 0.80, 0.70, 0.45, hence the
  default); on three one-minute conversations it found 3, 4 and 1 speakers
  for 3, 3 and 1. See docs/speakers.md.