Vendor wespeaker inference to eliminate pyannote Dockerfile stubs

[beastoin]

Context

PR #8082 added built-in wespeaker speaker embedding for batch diarization. PR #8085 patches the Dockerfile with 3 stubs to work around NGC ABI incompatibility with torchaudio/torch_audiomentations/opentelemetry.

Research by @geni (PR #8085 comment) confirmed:

• No NVIDIA-published torchaudio wheel exists for NGC (known issue since 2023)
• No pyannote inference-only install mode (torch_audiomentations is a hard dep)
• pyannote 4.0 adds torchcodec as hard dep (another C extension ABI crash) — <4.0 pin is critical
• All NGC users use the same stub workaround or avoid torchaudio entirely

Proposal

Vendor the wespeaker inference path (~60 lines) to eliminate pyannote.audio + ~25 transitive deps + all 3 Dockerfile stubs:

1. torchaudio.compliance.kaldi.fbank — pure Python, already working in NGC via the patched init
2. ResNet34 forward pass — load weights from HuggingFace, run inference directly
3. L2 normalize output → 256-dim embedding

This removes:

• pyannote.audio, pyannote.core, pyannote.database, pyannote.pipeline
• speechbrain, asteroid-filterbanks, einops, torch_audiomentations (stub)
• tensorboardX, hf_transfer, semver
• All 3 Dockerfile stubs (torchaudio patch, torch_audiomentations stub, telemetry stub)
• The <4.0 version pin (no longer needed)

Files

• backend/parakeet/transcribe.py:33-65 — current pyannote import + get_builtin_embedding_model()
• backend/parakeet/Dockerfile — current stubs

Acceptance criteria

• Vendored wespeaker inference produces identical 256-dim embeddings to pyannote Inference
• DER benchmark matches current implementation
• All 3 Dockerfile stubs removed
• pyannote.audio and transitive deps removed from Dockerfile
• Existing unit tests updated and passing
• Fallback to HTTP diarizer preserved