feat(transcription): GPU execution provider with CPU fallback (Track-B step 6)

[farce1]

Track-B step 6 — GPU execution provider with CPU fallback

WhisperEngine/ParakeetEngine previously hardcoded the ONNX Runtime provider
to "cpu". This selects a platform-appropriate execution provider instead —
CoreML on macOS, CPU elsewhere — and retries on CPU if a non-CPU provider
fails to load, so transcription can never regress relative to the CPU baseline.

Stacked on feat/asr-model-picker (#13): #9 ← #10 ← #11 ← #12 ← #13 ← this.

Design notes

• sherpa-rs is built with only download-binaries (no cuda feature), so the prebuilt binaries are CPU-only off macOS; CoreML is the only GPU EP this build can use. CUDA selection would be dead code, so it's intentionally omitted (YAGNI) until a CUDA-enabled build is introduced.
• select_provider(os) (pure) and with_cpu_fallback(provider, build) (the safety policy) are both unit-tested; load_with_provider carries the real recognizer config.
• Runtime GPU speedup is not verified in this environment (needs the models + audio); the tested CPU-fallback guarantees no functional regression regardless.

Gate

BE: clippy clean, cargo build ✓, cargo test --lib 12 passed (the only 3 failures are the inherited maintainer-owned REPLACE_WITH_ SHA256 release gate). FE: unchanged on this branch.

Merges once main's vendor + REPLACE_WITH_ blockers are cleared and the chain rebases.