Add Cohere Transcribe support; bump sherpa-onnx to v1.12.38 (0.7.0)

[1R053]

Summary

• New cohere_transcribe module wrapping the SherpaOnnxOfflineCohereTranscribeModelConfig C API added upstream in sherpa-onnx v1.12.x — 14-language ASR with native punctuation + ITN toggles.
• Bundled sherpa-onnx bumped v1.12.15 → v1.12.38 (submodule, dist.json, checksum.txt).
• First cargo-runnable integration tests in the repo (previously only examples/), with optional auto-download of models.

Breaking changes — 0.7.0

• ZipVoiceTtsConfig drops flow_matching_model / text_model / pinyin_dict (removed upstream) and adds encoder / decoder / lexicon (added upstream) to match the v1.12.38 C struct layout. Callers must rename field usage.
• Minimum bundled sherpa-onnx is v1.12.38.

Apple Silicon performance fix

dist.json for aarch64-apple-darwin now pulls sherpa-onnx-{tag}-onnxruntime-1.24.4-osx-arm64-shared.tar.bz2 (the full-optimization 35 MB onnxruntime build, bit-identical to the Python pip wheel's) instead of sherpa-onnx-{tag}-osx-arm64-shared.tar.bz2 (a smaller variant missing graph-optimization paths that blocked the post-first-inference kernel-cache warmup).

Net effect on spot-check: multi-x speedup on warm inferences, bringing warm-path parity with the Python pip wheel. Concretely, we saw Cohere Transcribe int8 warm-run RTF drop from ~0.13 to ~0.05 on Apple Silicon.

Also:

• Split the universal2 fat-binary entries into native single-arch tarballs for aarch64/x86_64 Darwin (smaller downloads, no lipo-imposed compiler compromises).
• Renamed the Windows asset to the new *-MD-Release.tar.bz2 upstream adopted.

Test suite

tests/offline_recognizers.rs + tests/test_utils.rs:

• Exercises both the updated Whisper path (regression guard for the v1.12.38 bump) and the new Cohere module end-to-end against real audio.
• ensure_model(&ModelArchive) / ensure_motivation_wav() helper resolves a cache dir from SHERPA_TEST_MODELS or workspace-root test_data/.
• Skips gracefully when files missing (CI-friendly default); auto-downloads when SHERPA_DOWNLOAD_MODELS=1.
• Downloads serialised via std::sync::Once for parallel-runner safety.
• Adding further model archives is a one-const entry.

Safe-field-addition refactor

Several recognizer constructors (moonshine, whisper, zipformer, transducer, sense_voice, paraformer, dolphin, tts/vits, tts/kokoro, tts/matcha, tts/kitten) switched their SherpaOnnxOfflineModelConfig / SherpaOnnxOfflineTtsModelConfig literals to ..Default::default() — covers the new upstream fields (cohere_transcribe, fire_red_asr_ctc, funasr_nano, merged_decoder, enable_segment_timestamps, enable_token_timestamps, pocket, supertonic) without field-by-field churn and keeps future additions no-break.

Test plan

• cargo test -p sherpa-rs --test offline_recognizers — skips cleanly without models (6 pass, no failures)
• SHERPA_DOWNLOAD_MODELS=1 cargo test -p sherpa-rs --test offline_recognizers --release -- --nocapture — downloads and runs whisper + cohere, both produce correct English transcripts (6/6 pass)
• Warm-run RTF matches Python sherpa_onnx pip wheel on Apple Silicon after the onnxruntime-1.24.4-osx-arm64-shared switch
• Non-macOS CI — Linux/Windows assets unchanged in mechanism, just renamed where upstream renamed them; needs validation by maintainers with those targets