fix(tts-engines): select device via torch.accelerator in confucius4 and dots_tts

Both Confucius4 and DOTS-TTS engine sidecars hardcode device selection to
`torch.cuda.is_available()`, which returns False on Ascend NPU, Intel XPU,
and other non-CUDA accelerators — causing the models to silently run on CPU
(in fp32) instead of the available accelerator.

Replace with `torch.accelerator.current_accelerator().type`, the
device-agnostic API that auto-detects CUDA, NPU, XPU, MPS, and CPU. MPS is
excluded for Confucius4 (upstream untested on Apple Silicon). dtype stays
bf16 for any GPU-class accelerator and fp32 on CPU.

Also verified on Ascend 910B (torch 2.14, torch_npu, 4 NPU):
  before: cuda_available=False → device "cpu", precision "float32"
  after:  accelerator → confucius4 device="npu", dots_tts precision="bfloat16"

backend/engines/confucius4/main.py

@@ -115,7 +115,9 @@ def _load_model(stdout):
     import torch
     from confuciustts.cli.inference import ConfuciusTTS  # type: ignore[import-not-found]
 
-    device = "cuda" if torch.cuda.is_available() else "cpu"
+    device = torch.accelerator.current_accelerator().type  # 'cuda', 'npu', 'mps', 'xpu', 'cpu'
+    if device == "mps":
+        device = "cpu"  # ConfuciusTTS is untested on MPS; fall back to CPU for safety
     _send(stdout, {"op": "progress", "stage": "loading_model", "percent": 50})
 
     _model = ConfuciusTTS(config_path=_config_path(), device=device)

backend/engines/dots_tts/main.py

@@ -115,7 +115,7 @@ def _load_runtime(stdout):
     from dots_tts.runtime import DotsTtsRuntime  # type: ignore[import-not-found]
 
     repo = os.environ.get("OMNIVOICE_DOTS_TTS_MODEL", _DEFAULT_REPO)
-    default_precision = "bfloat16" if torch.cuda.is_available() else "float32"
+    default_precision = "bfloat16" if torch.accelerator.current_accelerator().type != "cpu" else "float32"
     precision = os.environ.get("OMNIVOICE_DOTS_TTS_PRECISION", default_precision)
     optimize = os.environ.get("OMNIVOICE_DOTS_TTS_OPTIMIZE", "0") == "1"