Fix warnings in xcebra

[stes]

The current test pipeline gives the following error (e.g. this action run):

tests/test_attribution.py::test_neuron_gradient_method
tests/test_attribution.py::test_batched_methods
  /opt/hostedtoolcache/Python/3.10.19/x64/lib/python3.10/site-packages/captum/attr/_core/neuron/neuron_gradient.py:173: UserWarning: Input Tensor 0 did not already require gradients, required_grads has been set automatically.
    gradient_mask = apply_gradient_requirements(inputs_tuple)

tests/test_integration_xcebra.py::test_synthetic_data_training
  /home/runner/work/CEBRA/CEBRA/cebra/models/jacobian_regularizer.py:118: UserWarning: This overload of addcdiv is deprecated:
  	addcdiv(Tensor input, Number value, Tensor tensor1, Tensor tensor2, *, Tensor out = None)
  Consider using one of the following signatures instead:
  	addcdiv(Tensor input, Tensor tensor1, Tensor tensor2, *, Number value = 1, Tensor out = None) (Triggered internally at /pytorch/torch/csrc/utils/python_arg_parser.cpp:1805.)
    v = torch.addcdiv(arxilirary_zero, 1.0, v, vnorm)

[AmitSubhash]

Two default-visible warnings in the xCEBRA path, both fixable on our side:

1. A deprecated positional torch.addcdiv overload in jacobian_regularizer.py; the keyword form is bit-identical and silences it.
2. captum's "required_grads set automatically" from the two NeuronGradient methods in attribution_models.py (fires 25× in the attribution tests); setting requires_grad on the input before .attribute() clears it with identical outputs, and other attribution methods are unaffected.

The torch.jit.script deprecation is out of scope (already in #299). Happy to PR both plus a scoped warnings-as-errors guard if no one's on it.