feat: add SpeechBrain transcription provider

[franklincg]

Summary

Adds a local SpeechBrain ASR provider to Sapat's current provider registry.

• Uses speechbrain.inference.ASR.EncoderDecoderASR lazily, so the base package does not require SpeechBrain.
• Defaults to speechbrain/asr-crdnn-rnnlm-librispeech and supports SPEECHBRAIN_MODEL, SPEECHBRAIN_SAVEDIR, and SPEECHBRAIN_DEVICE.
• Adds the optional sapat[speechbrain] extra and keeps credentials unnecessary for local inference.
• Returns normalized TranscriptionResult text for common SpeechBrain output shapes.
• Adds mocked tests; CI does not download model weights.

Validation

• pytest tests/providers/test_speechbrain.py -q — 18 passed
• python -m compileall -q sapat tests — PASS
• black --check sapat/providers/speechbrain.py tests/providers/test_speechbrain.py — PASS
• git diff --check — PASS
• Full suite — 193 passed, 1 unrelated pre-existing Windows failure in TestWhisperXProvider::test_transcribe_builds_command_and_reads_output caused by existing WHISPERX_BINARY Windows path parsing.

No API keys, private audio, model files, or generated transcripts are committed.

Companion work for daytona/content#13 ($150 Algora bounty).