Local streaming speech-to-text on ESP32-P4 using Kroko model

An ESP32-P4 implementation achieves 4.34% WER on LibriSpeech test-clean using the 65.7M-parameter Kroko Zipformer2 transducer. The system runs entirely on-device without cloud or neural accelerators, utilizing 2-bit and 4-bit quantized weights to fit within 32 MB of PSRAM.