rai-perception only supports partial CPU fallback

[Juliaj]

Is your feature request related to a problem? Please describe.
Current rai-perception only supports partial CPU fallback:

• GDSegmenter — Has automatic CPU fallback:

   if use_cuda:       self.device = "cuda" if torch.cuda.is_available() else "cpu"

Falls back to CPU if CUDA isn't available.

• GDBoxer — No automatic fallback:

      if not use_cuda:
       self.model = Model(self.cfg_path, self.weight_path, device="cpu")
   else:
       self.model = Model(self.cfg_path, self.weight_path)  # Uses default (CUDA)

If use_cuda=True and CUDA isn't available, it may fail.

• Agents — Don't expose use_cuda; they always use the default (use_cuda=True).

Additional context
How important is it to support CPU fallback ?