Fix device for initial RHO-LOSS tensor (#525)

The initial tensor resides on CPU. For me, it failed when running
RHO-LOSS, since the very first ` self.rho_loss =
torch.cat([self.rho_loss, training_loss -
irreducible_loss]).to(training_loss.dtype)` was on two different
devices, the CPU (initial tensor) and the GPU (training loss and IR
loss). This PR fixes that by moving the initial tensor to the correct
device.

modyn/trainer_server/internal/trainer/remote_downsamplers/remote_rho_loss_downsampling.py

@@ -36,10 +36,11 @@ def __init__(
         )
         self.rho_loss: torch.Tensor = torch.tensor([])
         self.number_of_points_seen = 0
+        self._device = device
 
     def init_downsampler(self) -> None:
         self.index_sampleid_map: list[int] = []
-        self.rho_loss = torch.tensor([])
+        self.rho_loss = torch.tensor([]).to(self._device)
         self.number_of_points_seen = 0
 
     def inform_samples(