Add Argmax Sortformer

[EduardoPach]

What does this PR do?

• Integrate Sortformer as a new diarization model option alongside the existing pyannote3/pyannote4 clusterers across SpeakerKit, WhisperKitPro engine, and orchestration pipelines.
• Add new pipeline aliases for standalone Sortformer diarization (speakerkit-sortformer-compressed) and orchestration combos with compressed Parakeet v2/v3 models (whisperkitpro-orchestration-parakeet-v{2,3}-compressed-sortformer-compressed).
• Introduce diarization_mode config (realtime / prerecorded) to control Sortformer streaming latency, and a new subsegment orchestration strategy now used as the default for all WhisperKitPro orchestration pipelines.

Details

SpeakerKit diarization pipeline (speakerkit.py)

• Accept sortformer_model_name and sortformer_model_variant config fields with cross-validation.
• Move CLI arg construction and stdout parsing into SpeakerKitPipelineConfig so parsing logic adapts to the model type (Sortformer reports seconds; pyannote reports milliseconds).
• Skip num_speakers for Sortformer (unsupported) with a warning.

WhisperKitPro engine & orchestration (whisperkitpro_engine.py, orchestration_whisperkitpro.py)

• Extend clusterer_version literal to include sortformer and surface the new diarization_mode flag through CLI args.
• Fix config key from clusterer_version_string to clusterer_version.
• Enable --verbose by default on the engine CLI.

Pipeline aliases (pipeline_aliases.py)

• Register three new aliases for Sortformer-backed pipelines.
• Update all existing WhisperKitPro orchestration aliases to use subsegment strategy and the corrected clusterer_version key.

Test plan

• Run speakerkit-sortformer-compressed on a sample diarization dataset and verify RTTM output and reported prediction time.
• Run whisperkitpro-orchestration-parakeet-v3-compressed-sortformer-compressed end-to-end and confirm diarization segments are produced.
• Verify existing pyannote-based pipelines (speakerkit, whisperkitpro-orchestration-*) remain unaffected.
• Confirm diarization_mode flag is correctly forwarded (--diarization-mode prerecorded / realtime).

[arda-argmax]

LGTM!