Add AMD GPU support: faster_whisper CPU fallback + whisper.cpp backend

[Kaihui-AMD]

Summary

Makes WhisperLive work on AMD GPUs (#520):

1. Fix faster_whisper crashing on AMD/ROCm — fall back to CPU instead of crashing.
2. Add a whisper_cpp backend — real GPU transcription on AMD (Vulkan/ROCm), Intel, Apple.

Verified on a Radeon AI PRO R9700 (RDNA4, gfx1201) and a Ryzen AI Max+ 395 / Radeon 8060S
("Strix Halo", gfx1151 — the #520 chip). CUDA / TensorRT / OpenVINO paths unaffected.

1. faster_whisper: don't crash on AMD

CTranslate2's PyPI wheels are CUDA-only, but on ROCm torch.cuda.is_available() is True, so
the server picked device="cuda" and crashed (RuntimeError: CUDA failed ...). Guard on the
CTranslate2 device count instead:

use_cuda = torch.cuda.is_available() and ctranslate2.get_cuda_device_count() > 0
device = "cuda" if use_cuda else "cpu"

Applied to all three WhisperModel sites (the websocket backend + both REST endpoints in
server.py). CUDA users are unaffected; diarization/translation use torch and already work on
ROCm.

2. New whisper_cpp backend (GPU on AMD)

python run_server.py -b whisper_cpp

ServeClientWhisperCpp runs whisper.cpp via
pywhispercpp — ggml's Vulkan/ROCm/Metal backends
give real GPU transcription where CTranslate2 can't. Vulkan is the portable path (one build →
AMD/Intel/NVIDIA). It plugs into the existing server (same protocol/clients) and falls back to
faster_whisper if pywhispercpp is missing. Adds docker/Dockerfile.vulkan (turnkey) + README.

Build pywhispercpp with GPU (from git; setup.py forwards GGML_* to CMake):

GGML_VULKAN=1 pip install "git+https://github.com/absadiki/pywhispercpp.git"                    # portable
GGML_HIP=1 AMDGPU_TARGETS=gfx1151 pip install "git+https://github.com/absadiki/pywhispercpp.git" # ROCm

Verification

samples/jfk.wav (11 s), exact-text match; drove the real backend classes plus a full
run_server.py -b whisper_cpp end-to-end over WebSocket.

• faster_whisper fix: device=cuda crash → now CPU, transcribes correctly.
• whisper_cpp GPU backend:

box / chip	backend	device	correct?	time (cold/warm)¹
R9700 · gfx1201	Vulkan	Vulkan0 (RADV GFX1201)	✅	0.055 / 0.049 s
gfx1201 — full server E2E (WebSocket)	Vulkan	Vulkan0	✅	streamed OK
Ryzen AI Max+ 395 / 8060S · gfx1151 (#520 chip)	Vulkan	Vulkan0 (RADV GFX1151)	✅	0.336 / 0.071 s
gfx1151	ROCm/HIP	ROCm0	✅	0.135 / 0.077 s

¹ ggml-base.en, single runs — verification timings, not a formal benchmark.

Notes

• pywhispercpp must be source-built for GPU (PyPI is CPU-only); Dockerfile.vulkan handles it.
  RDNA4 needs a recent Mesa/RADV.
• First version: no_speech_prob is 0.0, word timestamps + whisper.cpp VAD not wired, and
  the OpenAI REST endpoint still uses CTranslate2 (CPU on AMD) — only the websocket backend is GPU.
• CTranslate2's ROCm forks are intentionally avoided (they produce wrong output on RDNA4).

[makaveli10]

Closing for #523