Add Nemotron ASR transcription support

[ncylich]

Summary

• add Nemotron ASR model-family detection, conversion, transpile, and runtime wiring
• add RNNT decoding support with file and stream transcription parity
• harden graph save/load parameter persistence for ASR-relevant ops
• restore corrected causal depthwise SIMD while preserving left-padded causal convolution semantics
• restore Core ML audio encoder fast paths and warm audio encoder variants used by file and stream transcription
• add focused regression tests for detection, graph serialization, and reused decoder state outputs

Benchmarks

Measured on cactus-engine/tests/assets/test.wav using mean engine time over 5 fresh CLI runs.

Model	Path	Mean engine time	Realtime factor	Speed vs CPU
Parakeet	CPU	738 ms	0.0368x	baseline
Parakeet	Apple/NPU	214 ms	0.0107x	3.45x faster
Nemotron	CPU	2178 ms	0.1087x	baseline
Nemotron	Apple/Core ML	586 ms	0.0293x	3.72x faster

Testing

• cmake --build cactus-kernels/build --target test_conv && ./cactus-kernels/build/test_conv
• source ./venv/bin/activate && cactus build && PYTHONPATH=python pytest python/cactus/convert/tests/test_policy.py::test_nemotron_asr_config_detection_extraction_and_plan python/cactus/convert/tests/test_policy.py::test_prompt_tdt_config_detection_stays_parakeet_tdt python/cactus/convert/tests/test_policy.py::test_nemotron_prompt_dim_preserves_sparse_prompt_ids python/tests/test_graph.py::TestGraphSaveLoad::test_save_load_roundtrip_glu_axis python/tests/test_transpile_aten_ops.py::test_reused_lstm_state_output_is_materialized -q
• source ./venv/bin/activate && cactus build && cactus transcribe nvidia/nemotron-3.5-asr-streaming-0.6b --file cactus-engine/tests/assets/test.wav
• source ./venv/bin/activate && cactus build && cactus test --component engine --suite stt --transcription-model nvidia/nemotron-3.5-asr-streaming-0.6b
• source ./venv/bin/activate && cactus build && cactus test --component engine --suite stream --transcription-model nvidia/nemotron-3.5-asr-streaming-0.6b
• source ./venv/bin/activate && cactus build && cactus test --component engine --suite stt --transcription-model /Users/noahcylich/Documents/Desert/cactus/weights/nemotron-3.5-asr-streaming-0.6b-cq4-apple
• source ./venv/bin/activate && cactus build && cactus test --component engine --suite stream --transcription-model /Users/noahcylich/Documents/Desert/cactus/weights/nemotron-3.5-asr-streaming-0.6b-cq4-apple
• source ./venv/bin/activate && cactus build && cactus test --component engine --suite stt --transcription-model nvidia/parakeet-tdt-0.6b-v3
• source ./venv/bin/activate && cactus build && cactus test --component engine --suite stream --transcription-model nvidia/parakeet-tdt-0.6b-v3
• source ./venv/bin/activate && cactus build && cactus test --component engine --suite stt --transcription-model /Users/noahcylich/Documents/Desert/cactus/weights/parakeet-tdt-0.6b-v3-cq4-apple
• source ./venv/bin/activate && cactus build && cactus test --component engine --suite stream --transcription-model /Users/noahcylich/Documents/Desert/cactus/weights/parakeet-tdt-0.6b-v3-cq4-apple
• source ./venv/bin/activate && cactus build && cactus test --component engine --suite stt --transcription-model openai/whisper-base
• source ./venv/bin/activate && cactus build && cactus test --component engine --suite stream --transcription-model openai/whisper-base
• git diff --check