Add Voxtral Small 24B as a transcription option

Whisper large-v3 mis-segments unfamiliar proper nouns: "Whisper Large"
came back as "Wisp Alarge", and it leaves doubled words and sentence
breaks in the wrong places. Voxtral decodes with a 24B language model,
so ambiguous acoustics get resolved against what a sentence can mean.
Checked both against saved captures: Voxtral fixed the proper noun and
the punctuation, at 1.7 s round-trip for 15 s of audio.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QrC2xo3RUExhVPzWGnGfRK

blurt.py

@@ -63,6 +63,11 @@ TRANSCRIPTION_MODELS = [
         "OpenRouter: gpt-4o-transcribe",
         "openrouter",
     ),
+    (
+        "mistralai/voxtral-small-24b-2507-stt",
+        "OpenRouter: voxtral-small-24b (LLM-based)",
+        "openrouter",
+    ),
 ]
 PROVIDER_KEY_HINTS = {
     "openai": ("set OPENAI_API_KEY or configure keyring/Credential Manager entry blurt/openai-api-key"),