feat: add tested native Transformers CUDA transcription recipe

[LauraGPT]

Summary

Add explicit CPU/CUDA and float32/BF16 selection to the native Transformers CLI while preserving CPU float32 defaults. Keep CUDA requirements separate from the CPU environment and document the tested hardware and functional scope.

Verification

• Red/green: two new contract tests failed before implementation; full workflow suite now 31 passed, rerun on this exact commit.
• H100 80 GB, Linux/Python 3.12, driver 550.127.08, Transformers 5.17.0, torch/torchaudio 2.11.0+cu128: eight native float32/BF16 English, Chinese, keywords and padded mixed-batch cases passed.
• Exact CLI: CUDA float32 English, CUDA BF16 Chinese/English batch, and original torch 2.10.0+cpu default English all completed with nonempty text and EOS.
• CUDA-hidden execution failed before sample/model download; no silent fallback.
• Independent read-only review found no P1/P2.

Boundaries and rollback

Public checkpoint revision d93b302ee7fd505e1b3576120fc142fc6f7820e1 only, trust_remote_code=False. Chinese recognition error remains even with keyword hint. Attention-dispatch warning retained; no accuracy, attention-kernel, throughput, minimum-VRAM, other-GPU or hosted-Space claim. No weights or shared environment changes.
Pre-push git bundle and exact source hashes saved on ind-gpu8; tests and raw outputs retained. Signed commit and DCO.