No device and batch size selection during eval

[abmazitov]

Current implementation of the mtt eval forces the model to run on the CPU only. This is caused by a recent metatensor/metatensor#668 PR, where the model is forced to be saved with a CPU device, and the following lines in the eval script

# Infer the device and dtype from the model
model_tensor = next(itertools.chain(model.parameters(), model.buffers()))
dtype = model_tensor.dtype
device = model_tensor.device

where device is inherited from the model parameters. Batch size selection is also not available, since it is hardcoded to be equal to 1.

# Create a dataloader
dataloader = torch.utils.data.DataLoader(
    dataset,
    batch_size=1,  # TODO: allow to set from outside!!
    collate_fn=collate_fn,
    shuffle=False,
)

Therefore, the evaluation part of the MTT is now extremely slow and rigid.