Use Parakeet as the default English model

[synesthesiam]

Summary

• replace the default English Citrinet model with Parakeet TDT-CTC 110M
• bundle the attributed Parakeet ONNX archive for offline first startup and pin the model dataset revision
• rename the generic ONNX acoustic backend to nemo, while accepting citrinet as a compatibility alias
• use the calibrated Parakeet max-score default of 3.8 and migrate persisted legacy English defaults of 5.0 without changing custom values
• cap NeMo ONNX inference at four threads
• align application help, documentation, benchmarks, and release metadata with the new model and score defaults

The model archive is published in rhasspy/rhasspy-speech at revision 9e6bcf3af07909f39f71674f62372508d00dd4a9. It includes provenance and NVIDIA CC-BY-4.0 attribution.

Validation

• focused release, settings, numeric-range, and debug-mode tests
• formatting, type checking, and Python compilation for the changed runtime and evaluation tools
• English sentence-FST evaluation: 46/46 standard commands
• Voice Preview Edition corpus: 62/67 commands at the zero-false-accept gate