Audio fixtures for the gemma-4 speech path (LLM/tests/Gemma4AudioTests.swift).

The SAME sentence in two languages, read by native speakers, so a
transcription test measures the model rather than the fixture. Both come
from Tatoeba (https://tatoeba.org) under CC BY 2.0 FR.

  en-lets-leave-tomorrow.wav
      "Let's leave tomorrow."
      sentence #6066996 by CK, audio #644931 by CK

  fr-partons-demain.wav
      "Partons demain."
      sentence #11460054 by lbdx, audio #1102917 by Adrien_FR

Licence: CC BY 2.0 FR -- https://creativecommons.org/licenses/by/2.0/fr/
Converted from the source MP3 to 16 kHz mono WAV (the rate
gemma4.audio.mel.sample_rate names); no other processing.

WHY these and not shorter clips: the model needs roughly a second of speech.
Measured on single words, a 0.60s clip (15 soft tokens) returns "no
discernible spoken words", while 0.77s (19) and 1.10s (28) transcribe
correctly. These are 1.67s and 1.41s.
